Line data Source code
1 : //! Timeline repository implementation that keeps old data in layer files, and
2 : //! the recent changes in ephemeral files.
3 : //!
4 : //! See tenant/*_layer.rs files. The functions here are responsible for locating
5 : //! the correct layer for the get/put call, walking back the timeline branching
6 : //! history as needed.
7 : //!
8 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
9 : //! directory. See docs/pageserver-storage.md for how the files are managed.
10 : //! In addition to the layer files, there is a metadata file in the same
11 : //! directory that contains information about the timeline, in particular its
12 : //! parent timeline, and the last LSN that has been written to disk.
13 : //!
14 :
15 : use std::collections::hash_map::Entry;
16 : use std::collections::{BTreeMap, HashMap, HashSet};
17 : use std::fmt::{Debug, Display};
18 : use std::fs::File;
19 : use std::future::Future;
20 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
21 : use std::sync::{Arc, Mutex, Weak};
22 : use std::time::{Duration, Instant, SystemTime};
23 : use std::{fmt, fs};
24 :
25 : use anyhow::{Context, bail};
26 : use arc_swap::ArcSwap;
27 : use camino::{Utf8Path, Utf8PathBuf};
28 : use chrono::NaiveDateTime;
29 : use enumset::EnumSet;
30 : use futures::StreamExt;
31 : use futures::stream::FuturesUnordered;
32 : use itertools::Itertools as _;
33 : use once_cell::sync::Lazy;
34 : pub use pageserver_api::models::TenantState;
35 : use pageserver_api::models::{self, RelSizeMigration};
36 : use pageserver_api::models::{
37 : CompactInfoResponse, TimelineArchivalState, TimelineState, TopTenantShardItem,
38 : WalRedoManagerStatus,
39 : };
40 : use pageserver_api::shard::{ShardIdentity, ShardStripeSize, TenantShardId};
41 : use postgres_ffi::PgMajorVersion;
42 : use remote_storage::{DownloadError, GenericRemoteStorage, TimeoutOrCancel};
43 : use remote_timeline_client::index::GcCompactionState;
44 : use remote_timeline_client::manifest::{
45 : LATEST_TENANT_MANIFEST_VERSION, OffloadedTimelineManifest, TenantManifest,
46 : };
47 : use remote_timeline_client::{
48 : FAILED_REMOTE_OP_RETRIES, FAILED_UPLOAD_WARN_THRESHOLD, UploadQueueNotReadyError,
49 : download_tenant_manifest,
50 : };
51 : use secondary::heatmap::{HeatMapTenant, HeatMapTimeline};
52 : use storage_broker::BrokerClientChannel;
53 : use timeline::compaction::{CompactionOutcome, GcCompactionQueue};
54 : use timeline::import_pgdata::ImportingTimeline;
55 : use timeline::layer_manager::LayerManagerLockHolder;
56 : use timeline::offload::{OffloadError, offload_timeline};
57 : use timeline::{
58 : CompactFlags, CompactOptions, CompactionError, PreviousHeatmap, ShutdownMode, import_pgdata,
59 : };
60 : use tokio::io::BufReader;
61 : use tokio::sync::{Notify, Semaphore, watch};
62 : use tokio::task::JoinSet;
63 : use tokio_util::sync::CancellationToken;
64 : use tracing::*;
65 : use upload_queue::NotInitialized;
66 : use utils::circuit_breaker::CircuitBreaker;
67 : use utils::crashsafe::path_with_suffix_extension;
68 : use utils::sync::gate::{Gate, GateGuard};
69 : use utils::timeout::{TimeoutCancellableError, timeout_cancellable};
70 : use utils::try_rcu::ArcSwapExt;
71 : use utils::zstd::{create_zst_tarball, extract_zst_tarball};
72 : use utils::{backoff, completion, failpoint_support, fs_ext, pausable_failpoint};
73 :
74 : use self::config::{AttachedLocationConfig, AttachmentMode, LocationConf};
75 : use self::metadata::TimelineMetadata;
76 : use self::mgr::{GetActiveTenantError, GetTenantError};
77 : use self::remote_timeline_client::upload::{upload_index_part, upload_tenant_manifest};
78 : use self::remote_timeline_client::{RemoteTimelineClient, WaitCompletionError};
79 : use self::timeline::uninit::{TimelineCreateGuard, TimelineExclusionError, UninitializedTimeline};
80 : use self::timeline::{
81 : EvictionTaskTenantState, GcCutoffs, TimelineDeleteProgress, TimelineResources, WaitLsnError,
82 : };
83 : use crate::basebackup_cache::BasebackupCache;
84 : use crate::config::PageServerConf;
85 : use crate::context;
86 : use crate::context::RequestContextBuilder;
87 : use crate::context::{DownloadBehavior, RequestContext};
88 : use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
89 : use crate::feature_resolver::{FeatureResolver, TenantFeatureResolver};
90 : use crate::l0_flush::L0FlushGlobalState;
91 : use crate::metrics::{
92 : BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN, CONCURRENT_INITDBS,
93 : INITDB_RUN_TIME, INITDB_SEMAPHORE_ACQUISITION_TIME, TENANT, TENANT_OFFLOADED_TIMELINES,
94 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC, TIMELINE_STATE_METRIC,
95 : remove_tenant_metrics,
96 : };
97 : use crate::task_mgr::TaskKind;
98 : use crate::tenant::config::LocationMode;
99 : use crate::tenant::gc_result::GcResult;
100 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
101 : use crate::tenant::remote_timeline_client::{
102 : INITDB_PATH, MaybeDeletedIndexPart, remote_initdb_archive_path,
103 : };
104 : use crate::tenant::storage_layer::{DeltaLayer, ImageLayer};
105 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
106 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
107 : use crate::virtual_file::VirtualFile;
108 : use crate::walingest::WalLagCooldown;
109 : use crate::walredo::{PostgresRedoManager, RedoAttemptType};
110 : use crate::{InitializationOrder, TEMP_FILE_SUFFIX, import_datadir, span, task_mgr, walredo};
111 :
112 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
113 : use utils::crashsafe;
114 : use utils::generation::Generation;
115 : use utils::id::TimelineId;
116 : use utils::lsn::{Lsn, RecordLsn};
117 :
118 : pub mod blob_io;
119 : pub mod block_io;
120 : pub mod vectored_blob_io;
121 :
122 : pub mod disk_btree;
123 : pub(crate) mod ephemeral_file;
124 : pub mod layer_map;
125 :
126 : pub mod metadata;
127 : pub mod remote_timeline_client;
128 : pub mod storage_layer;
129 :
130 : pub mod checks;
131 : pub mod config;
132 : pub mod mgr;
133 : pub mod secondary;
134 : pub mod tasks;
135 : pub mod upload_queue;
136 :
137 : pub(crate) mod timeline;
138 :
139 : pub mod size;
140 :
141 : mod gc_block;
142 : mod gc_result;
143 : pub(crate) mod throttle;
144 :
145 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
146 :
147 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
148 : // re-export for use in walreceiver
149 : pub use crate::tenant::timeline::WalReceiverInfo;
150 :
151 : /// The "tenants" part of `tenants/<tenant>/timelines...`
152 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
153 :
154 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
155 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
156 :
157 : /// References to shared objects that are passed into each tenant, such
158 : /// as the shared remote storage client and process initialization state.
159 : #[derive(Clone)]
160 : pub struct TenantSharedResources {
161 : pub broker_client: storage_broker::BrokerClientChannel,
162 : pub remote_storage: GenericRemoteStorage,
163 : pub deletion_queue_client: DeletionQueueClient,
164 : pub l0_flush_global_state: L0FlushGlobalState,
165 : pub basebackup_cache: Arc<BasebackupCache>,
166 : pub feature_resolver: FeatureResolver,
167 : }
168 :
169 : /// A [`TenantShard`] is really an _attached_ tenant. The configuration
170 : /// for an attached tenant is a subset of the [`LocationConf`], represented
171 : /// in this struct.
172 : #[derive(Clone)]
173 : pub(super) struct AttachedTenantConf {
174 : tenant_conf: pageserver_api::models::TenantConfig,
175 : location: AttachedLocationConfig,
176 : /// The deadline before which we are blocked from GC so that
177 : /// leases have a chance to be renewed.
178 : lsn_lease_deadline: Option<tokio::time::Instant>,
179 : }
180 :
181 : impl AttachedTenantConf {
182 118 : fn new(
183 118 : conf: &'static PageServerConf,
184 118 : tenant_conf: pageserver_api::models::TenantConfig,
185 118 : location: AttachedLocationConfig,
186 118 : ) -> Self {
187 : // Sets a deadline before which we cannot proceed to GC due to lsn lease.
188 : //
189 : // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
190 : // length, we guarantee that all the leases we granted before will have a chance to renew
191 : // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
192 118 : let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
193 118 : Some(
194 118 : tokio::time::Instant::now()
195 118 : + TenantShard::get_lsn_lease_length_impl(conf, &tenant_conf),
196 118 : )
197 : } else {
198 : // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
199 : // because we don't do GC in these modes.
200 0 : None
201 : };
202 :
203 118 : Self {
204 118 : tenant_conf,
205 118 : location,
206 118 : lsn_lease_deadline,
207 118 : }
208 118 : }
209 :
210 118 : fn try_from(
211 118 : conf: &'static PageServerConf,
212 118 : location_conf: LocationConf,
213 118 : ) -> anyhow::Result<Self> {
214 118 : match &location_conf.mode {
215 118 : LocationMode::Attached(attach_conf) => {
216 118 : Ok(Self::new(conf, location_conf.tenant_conf, *attach_conf))
217 : }
218 : LocationMode::Secondary(_) => {
219 0 : anyhow::bail!(
220 0 : "Attempted to construct AttachedTenantConf from a LocationConf in secondary mode"
221 : )
222 : }
223 : }
224 118 : }
225 :
226 381 : fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
227 381 : self.lsn_lease_deadline
228 381 : .map(|d| tokio::time::Instant::now() < d)
229 381 : .unwrap_or(false)
230 381 : }
231 : }
232 : struct TimelinePreload {
233 : timeline_id: TimelineId,
234 : client: RemoteTimelineClient,
235 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
236 : previous_heatmap: Option<PreviousHeatmap>,
237 : }
238 :
239 : pub(crate) struct TenantPreload {
240 : /// The tenant manifest from remote storage, or None if no manifest was found.
241 : tenant_manifest: Option<TenantManifest>,
242 : /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
243 : timelines: HashMap<TimelineId, Option<TimelinePreload>>,
244 : }
245 :
246 : /// When we spawn a tenant, there is a special mode for tenant creation that
247 : /// avoids trying to read anything from remote storage.
248 : pub(crate) enum SpawnMode {
249 : /// Activate as soon as possible
250 : Eager,
251 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
252 : Lazy,
253 : }
254 :
255 : ///
256 : /// Tenant consists of multiple timelines. Keep them in a hash table.
257 : ///
258 : pub struct TenantShard {
259 : // Global pageserver config parameters
260 : pub conf: &'static PageServerConf,
261 :
262 : /// The value creation timestamp, used to measure activation delay, see:
263 : /// <https://github.com/neondatabase/neon/issues/4025>
264 : constructed_at: Instant,
265 :
266 : state: watch::Sender<TenantState>,
267 :
268 : // Overridden tenant-specific config parameters.
269 : // We keep pageserver_api::models::TenantConfig sturct here to preserve the information
270 : // about parameters that are not set.
271 : // This is necessary to allow global config updates.
272 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
273 :
274 : tenant_shard_id: TenantShardId,
275 :
276 : // The detailed sharding information, beyond the number/count in tenant_shard_id
277 : shard_identity: ShardIdentity,
278 :
279 : /// The remote storage generation, used to protect S3 objects from split-brain.
280 : /// Does not change over the lifetime of the [`TenantShard`] object.
281 : ///
282 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
283 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
284 : generation: Generation,
285 :
286 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
287 :
288 : /// During timeline creation, we first insert the TimelineId to the
289 : /// creating map, then `timelines`, then remove it from the creating map.
290 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
291 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
292 :
293 : /// Possibly offloaded and archived timelines
294 : /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
295 : timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
296 :
297 : /// Tracks the timelines that are currently importing into this tenant shard.
298 : ///
299 : /// Note that importing timelines are also present in [`Self::timelines_creating`].
300 : /// Keep this in mind when ordering lock acquisition.
301 : ///
302 : /// Lifetime:
303 : /// * An imported timeline is created while scanning the bucket on tenant attach
304 : /// if the index part contains an `import_pgdata` entry and said field marks the import
305 : /// as in progress.
306 : /// * Imported timelines are removed when the storage controller calls the post timeline
307 : /// import activation endpoint.
308 : timelines_importing: std::sync::Mutex<HashMap<TimelineId, Arc<ImportingTimeline>>>,
309 :
310 : /// The last tenant manifest known to be in remote storage. None if the manifest has not yet
311 : /// been either downloaded or uploaded. Always Some after tenant attach.
312 : ///
313 : /// Initially populated during tenant attach, updated via `maybe_upload_tenant_manifest`.
314 : ///
315 : /// Do not modify this directly. It is used to check whether a new manifest needs to be
316 : /// uploaded. The manifest is constructed in `build_tenant_manifest`, and uploaded via
317 : /// `maybe_upload_tenant_manifest`.
318 : remote_tenant_manifest: tokio::sync::Mutex<Option<TenantManifest>>,
319 :
320 : // This mutex prevents creation of new timelines during GC.
321 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
322 : // `timelines` mutex during all GC iteration
323 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
324 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
325 : // timeout...
326 : gc_cs: tokio::sync::Mutex<()>,
327 : walredo_mgr: Option<Arc<WalRedoManager>>,
328 :
329 : /// Provides access to timeline data sitting in the remote storage.
330 : pub(crate) remote_storage: GenericRemoteStorage,
331 :
332 : /// Access to global deletion queue for when this tenant wants to schedule a deletion.
333 : deletion_queue_client: DeletionQueueClient,
334 :
335 : /// A channel to send async requests to prepare a basebackup for the basebackup cache.
336 : basebackup_cache: Arc<BasebackupCache>,
337 :
338 : /// Cached logical sizes updated updated on each [`TenantShard::gather_size_inputs`].
339 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
340 : cached_synthetic_tenant_size: Arc<AtomicU64>,
341 :
342 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
343 :
344 : /// Track repeated failures to compact, so that we can back off.
345 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
346 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
347 :
348 : /// Signals the tenant compaction loop that there is L0 compaction work to be done.
349 : pub(crate) l0_compaction_trigger: Arc<Notify>,
350 :
351 : /// Scheduled gc-compaction tasks.
352 : scheduled_compaction_tasks: std::sync::Mutex<HashMap<TimelineId, Arc<GcCompactionQueue>>>,
353 :
354 : /// If the tenant is in Activating state, notify this to encourage it
355 : /// to proceed to Active as soon as possible, rather than waiting for lazy
356 : /// background warmup.
357 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
358 :
359 : /// Time it took for the tenant to activate. Zero if not active yet.
360 : attach_wal_lag_cooldown: Arc<std::sync::OnceLock<WalLagCooldown>>,
361 :
362 : // Cancellation token fires when we have entered shutdown(). This is a parent of
363 : // Timelines' cancellation token.
364 : pub(crate) cancel: CancellationToken,
365 :
366 : // Users of the TenantShard such as the page service must take this Gate to avoid
367 : // trying to use a TenantShard which is shutting down.
368 : pub(crate) gate: Gate,
369 :
370 : /// Throttle applied at the top of [`Timeline::get`].
371 : /// All [`TenantShard::timelines`] of a given [`TenantShard`] instance share the same [`throttle::Throttle`] instance.
372 : pub(crate) pagestream_throttle: Arc<throttle::Throttle>,
373 :
374 : pub(crate) pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
375 :
376 : /// An ongoing timeline detach concurrency limiter.
377 : ///
378 : /// As a tenant will likely be restarted as part of timeline detach ancestor it makes no sense
379 : /// to have two running at the same time. A different one can be started if an earlier one
380 : /// has failed for whatever reason.
381 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
382 :
383 : /// `index_part.json` based gc blocking reason tracking.
384 : ///
385 : /// New gc iterations must start a new iteration by acquiring `GcBlock::start` before
386 : /// proceeding.
387 : pub(crate) gc_block: gc_block::GcBlock,
388 :
389 : l0_flush_global_state: L0FlushGlobalState,
390 :
391 : pub(crate) feature_resolver: Arc<TenantFeatureResolver>,
392 : }
393 : impl std::fmt::Debug for TenantShard {
394 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
395 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
396 0 : }
397 : }
398 :
399 : pub(crate) enum WalRedoManager {
400 : Prod(WalredoManagerId, PostgresRedoManager),
401 : #[cfg(test)]
402 : Test(harness::TestRedoManager),
403 : }
404 :
405 : #[derive(thiserror::Error, Debug)]
406 : #[error("pageserver is shutting down")]
407 : pub(crate) struct GlobalShutDown;
408 :
409 : impl WalRedoManager {
410 0 : pub(crate) fn new(mgr: PostgresRedoManager) -> Result<Arc<Self>, GlobalShutDown> {
411 0 : let id = WalredoManagerId::next();
412 0 : let arc = Arc::new(Self::Prod(id, mgr));
413 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
414 0 : match &mut *guard {
415 0 : Some(map) => {
416 0 : map.insert(id, Arc::downgrade(&arc));
417 0 : Ok(arc)
418 : }
419 0 : None => Err(GlobalShutDown),
420 : }
421 0 : }
422 : }
423 :
424 : impl Drop for WalRedoManager {
425 5 : fn drop(&mut self) {
426 5 : match self {
427 0 : Self::Prod(id, _) => {
428 0 : let mut guard = WALREDO_MANAGERS.lock().unwrap();
429 0 : if let Some(map) = &mut *guard {
430 0 : map.remove(id).expect("new() registers, drop() unregisters");
431 0 : }
432 : }
433 : #[cfg(test)]
434 5 : Self::Test(_) => {
435 5 : // Not applicable to test redo manager
436 5 : }
437 : }
438 5 : }
439 : }
440 :
441 : /// Global registry of all walredo managers so that [`crate::shutdown_pageserver`] can shut down
442 : /// the walredo processes outside of the regular order.
443 : ///
444 : /// This is necessary to work around a systemd bug where it freezes if there are
445 : /// walredo processes left => <https://github.com/neondatabase/cloud/issues/11387>
446 : #[allow(clippy::type_complexity)]
447 : pub(crate) static WALREDO_MANAGERS: once_cell::sync::Lazy<
448 : Mutex<Option<HashMap<WalredoManagerId, Weak<WalRedoManager>>>>,
449 0 : > = once_cell::sync::Lazy::new(|| Mutex::new(Some(HashMap::new())));
450 : #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
451 : pub(crate) struct WalredoManagerId(u64);
452 : impl WalredoManagerId {
453 0 : pub fn next() -> Self {
454 : static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
455 0 : let id = NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
456 0 : if id == 0 {
457 0 : panic!(
458 0 : "WalredoManagerId::new() returned 0, indicating wraparound, risking it's no longer unique"
459 : );
460 0 : }
461 0 : Self(id)
462 0 : }
463 : }
464 :
465 : #[cfg(test)]
466 : impl From<harness::TestRedoManager> for WalRedoManager {
467 118 : fn from(mgr: harness::TestRedoManager) -> Self {
468 118 : Self::Test(mgr)
469 118 : }
470 : }
471 :
472 : impl WalRedoManager {
473 3 : pub(crate) async fn shutdown(&self) -> bool {
474 3 : match self {
475 0 : Self::Prod(_, mgr) => mgr.shutdown().await,
476 : #[cfg(test)]
477 : Self::Test(_) => {
478 : // Not applicable to test redo manager
479 3 : true
480 : }
481 : }
482 3 : }
483 :
484 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
485 0 : match self {
486 0 : Self::Prod(_, mgr) => mgr.maybe_quiesce(idle_timeout),
487 : #[cfg(test)]
488 0 : Self::Test(_) => {
489 0 : // Not applicable to test redo manager
490 0 : }
491 : }
492 0 : }
493 :
494 : /// # Cancel-Safety
495 : ///
496 : /// This method is cancellation-safe.
497 26774 : pub async fn request_redo(
498 26774 : &self,
499 26774 : key: pageserver_api::key::Key,
500 26774 : lsn: Lsn,
501 26774 : base_img: Option<(Lsn, bytes::Bytes)>,
502 26774 : records: Vec<(Lsn, wal_decoder::models::record::NeonWalRecord)>,
503 26774 : pg_version: PgMajorVersion,
504 26774 : redo_attempt_type: RedoAttemptType,
505 26774 : ) -> Result<bytes::Bytes, walredo::Error> {
506 26774 : match self {
507 0 : Self::Prod(_, mgr) => {
508 0 : mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
509 0 : .await
510 : }
511 : #[cfg(test)]
512 26774 : Self::Test(mgr) => {
513 26774 : mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
514 26774 : .await
515 : }
516 : }
517 26774 : }
518 :
519 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
520 0 : match self {
521 0 : WalRedoManager::Prod(_, m) => Some(m.status()),
522 : #[cfg(test)]
523 0 : WalRedoManager::Test(_) => None,
524 : }
525 0 : }
526 : }
527 :
528 : /// A very lightweight memory representation of an offloaded timeline.
529 : ///
530 : /// We need to store the list of offloaded timelines so that we can perform operations on them,
531 : /// like unoffloading them, or (at a later date), decide to perform flattening.
532 : /// This type has a much smaller memory impact than [`Timeline`], and thus we can store many
533 : /// more offloaded timelines than we can manage ones that aren't.
534 : pub struct OffloadedTimeline {
535 : pub tenant_shard_id: TenantShardId,
536 : pub timeline_id: TimelineId,
537 : pub ancestor_timeline_id: Option<TimelineId>,
538 : /// Whether to retain the branch lsn at the ancestor or not
539 : pub ancestor_retain_lsn: Option<Lsn>,
540 :
541 : /// When the timeline was archived.
542 : ///
543 : /// Present for future flattening deliberations.
544 : pub archived_at: NaiveDateTime,
545 :
546 : /// Prevent two tasks from deleting the timeline at the same time. If held, the
547 : /// timeline is being deleted. If 'true', the timeline has already been deleted.
548 : pub delete_progress: TimelineDeleteProgress,
549 :
550 : /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
551 : pub deleted_from_ancestor: AtomicBool,
552 :
553 : _metrics_guard: OffloadedTimelineMetricsGuard,
554 : }
555 :
556 : /// Increases the offloaded timeline count metric when created, and decreases when dropped.
557 : struct OffloadedTimelineMetricsGuard;
558 :
559 : impl OffloadedTimelineMetricsGuard {
560 1 : fn new() -> Self {
561 1 : TIMELINE_STATE_METRIC
562 1 : .with_label_values(&["offloaded"])
563 1 : .inc();
564 1 : Self
565 1 : }
566 : }
567 :
568 : impl Drop for OffloadedTimelineMetricsGuard {
569 1 : fn drop(&mut self) {
570 1 : TIMELINE_STATE_METRIC
571 1 : .with_label_values(&["offloaded"])
572 1 : .dec();
573 1 : }
574 : }
575 :
576 : impl OffloadedTimeline {
577 : /// Obtains an offloaded timeline from a given timeline object.
578 : ///
579 : /// Returns `None` if the `archived_at` flag couldn't be obtained, i.e.
580 : /// the timeline is not in a stopped state.
581 : /// Panics if the timeline is not archived.
582 1 : fn from_timeline(timeline: &Timeline) -> Result<Self, UploadQueueNotReadyError> {
583 1 : let (ancestor_retain_lsn, ancestor_timeline_id) =
584 1 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
585 1 : let ancestor_lsn = timeline.get_ancestor_lsn();
586 1 : let ancestor_timeline_id = ancestor_timeline.timeline_id;
587 1 : let mut gc_info = ancestor_timeline.gc_info.write().unwrap();
588 1 : gc_info.insert_child(timeline.timeline_id, ancestor_lsn, MaybeOffloaded::Yes);
589 1 : (Some(ancestor_lsn), Some(ancestor_timeline_id))
590 : } else {
591 0 : (None, None)
592 : };
593 1 : let archived_at = timeline
594 1 : .remote_client
595 1 : .archived_at_stopped_queue()?
596 1 : .expect("must be called on an archived timeline");
597 1 : Ok(Self {
598 1 : tenant_shard_id: timeline.tenant_shard_id,
599 1 : timeline_id: timeline.timeline_id,
600 1 : ancestor_timeline_id,
601 1 : ancestor_retain_lsn,
602 1 : archived_at,
603 1 :
604 1 : delete_progress: timeline.delete_progress.clone(),
605 1 : deleted_from_ancestor: AtomicBool::new(false),
606 1 :
607 1 : _metrics_guard: OffloadedTimelineMetricsGuard::new(),
608 1 : })
609 1 : }
610 0 : fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
611 : // We expect to reach this case in tenant loading, where the `retain_lsn` is populated in the parent's `gc_info`
612 : // by the `initialize_gc_info` function.
613 : let OffloadedTimelineManifest {
614 0 : timeline_id,
615 0 : ancestor_timeline_id,
616 0 : ancestor_retain_lsn,
617 0 : archived_at,
618 0 : } = *manifest;
619 0 : Self {
620 0 : tenant_shard_id,
621 0 : timeline_id,
622 0 : ancestor_timeline_id,
623 0 : ancestor_retain_lsn,
624 0 : archived_at,
625 0 : delete_progress: TimelineDeleteProgress::default(),
626 0 : deleted_from_ancestor: AtomicBool::new(false),
627 0 : _metrics_guard: OffloadedTimelineMetricsGuard::new(),
628 0 : }
629 0 : }
630 1 : fn manifest(&self) -> OffloadedTimelineManifest {
631 : let Self {
632 1 : timeline_id,
633 1 : ancestor_timeline_id,
634 1 : ancestor_retain_lsn,
635 1 : archived_at,
636 : ..
637 1 : } = self;
638 1 : OffloadedTimelineManifest {
639 1 : timeline_id: *timeline_id,
640 1 : ancestor_timeline_id: *ancestor_timeline_id,
641 1 : ancestor_retain_lsn: *ancestor_retain_lsn,
642 1 : archived_at: *archived_at,
643 1 : }
644 1 : }
645 : /// Delete this timeline's retain_lsn from its ancestor, if present in the given tenant
646 0 : fn delete_from_ancestor_with_timelines(
647 0 : &self,
648 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
649 0 : ) {
650 0 : if let (Some(_retain_lsn), Some(ancestor_timeline_id)) =
651 0 : (self.ancestor_retain_lsn, self.ancestor_timeline_id)
652 : {
653 0 : if let Some((_, ancestor_timeline)) = timelines
654 0 : .iter()
655 0 : .find(|(tid, _tl)| **tid == ancestor_timeline_id)
656 : {
657 0 : let removal_happened = ancestor_timeline
658 0 : .gc_info
659 0 : .write()
660 0 : .unwrap()
661 0 : .remove_child_offloaded(self.timeline_id);
662 0 : if !removal_happened {
663 0 : tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
664 0 : "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
665 0 : }
666 0 : }
667 0 : }
668 0 : self.deleted_from_ancestor.store(true, Ordering::Release);
669 0 : }
670 : /// Call [`Self::delete_from_ancestor_with_timelines`] instead if possible.
671 : ///
672 : /// As the entire tenant is being dropped, don't bother deregistering the `retain_lsn` from the ancestor.
673 1 : fn defuse_for_tenant_drop(&self) {
674 1 : self.deleted_from_ancestor.store(true, Ordering::Release);
675 1 : }
676 : }
677 :
678 : impl fmt::Debug for OffloadedTimeline {
679 0 : fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
680 0 : write!(f, "OffloadedTimeline<{}>", self.timeline_id)
681 0 : }
682 : }
683 :
684 : impl Drop for OffloadedTimeline {
685 1 : fn drop(&mut self) {
686 1 : if !self.deleted_from_ancestor.load(Ordering::Acquire) {
687 0 : tracing::warn!(
688 0 : "offloaded timeline {} was dropped without having cleaned it up at the ancestor",
689 : self.timeline_id
690 : );
691 1 : }
692 1 : }
693 : }
694 :
695 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
696 : pub enum MaybeOffloaded {
697 : Yes,
698 : No,
699 : }
700 :
701 : #[derive(Clone, Debug)]
702 : pub enum TimelineOrOffloaded {
703 : Timeline(Arc<Timeline>),
704 : Offloaded(Arc<OffloadedTimeline>),
705 : Importing(Arc<ImportingTimeline>),
706 : }
707 :
708 : impl TimelineOrOffloaded {
709 0 : pub fn arc_ref(&self) -> TimelineOrOffloadedArcRef<'_> {
710 0 : match self {
711 0 : TimelineOrOffloaded::Timeline(timeline) => {
712 0 : TimelineOrOffloadedArcRef::Timeline(timeline)
713 : }
714 0 : TimelineOrOffloaded::Offloaded(offloaded) => {
715 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded)
716 : }
717 0 : TimelineOrOffloaded::Importing(importing) => {
718 0 : TimelineOrOffloadedArcRef::Importing(importing)
719 : }
720 : }
721 0 : }
722 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
723 0 : self.arc_ref().tenant_shard_id()
724 0 : }
725 0 : pub fn timeline_id(&self) -> TimelineId {
726 0 : self.arc_ref().timeline_id()
727 0 : }
728 1 : pub fn delete_progress(&self) -> &Arc<tokio::sync::Mutex<DeleteTimelineFlow>> {
729 1 : match self {
730 1 : TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
731 0 : TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
732 0 : TimelineOrOffloaded::Importing(importing) => &importing.delete_progress,
733 : }
734 1 : }
735 0 : fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
736 0 : match self {
737 0 : TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
738 0 : TimelineOrOffloaded::Offloaded(_offloaded) => None,
739 0 : TimelineOrOffloaded::Importing(importing) => {
740 0 : Some(importing.timeline.remote_client.clone())
741 : }
742 : }
743 0 : }
744 : }
745 :
746 : pub enum TimelineOrOffloadedArcRef<'a> {
747 : Timeline(&'a Arc<Timeline>),
748 : Offloaded(&'a Arc<OffloadedTimeline>),
749 : Importing(&'a Arc<ImportingTimeline>),
750 : }
751 :
752 : impl TimelineOrOffloadedArcRef<'_> {
753 0 : pub fn tenant_shard_id(&self) -> TenantShardId {
754 0 : match self {
755 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
756 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
757 0 : TimelineOrOffloadedArcRef::Importing(importing) => importing.timeline.tenant_shard_id,
758 : }
759 0 : }
760 0 : pub fn timeline_id(&self) -> TimelineId {
761 0 : match self {
762 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
763 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
764 0 : TimelineOrOffloadedArcRef::Importing(importing) => importing.timeline.timeline_id,
765 : }
766 0 : }
767 : }
768 :
769 : impl<'a> From<&'a Arc<Timeline>> for TimelineOrOffloadedArcRef<'a> {
770 0 : fn from(timeline: &'a Arc<Timeline>) -> Self {
771 0 : Self::Timeline(timeline)
772 0 : }
773 : }
774 :
775 : impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
776 0 : fn from(timeline: &'a Arc<OffloadedTimeline>) -> Self {
777 0 : Self::Offloaded(timeline)
778 0 : }
779 : }
780 :
781 : impl<'a> From<&'a Arc<ImportingTimeline>> for TimelineOrOffloadedArcRef<'a> {
782 0 : fn from(timeline: &'a Arc<ImportingTimeline>) -> Self {
783 0 : Self::Importing(timeline)
784 0 : }
785 : }
786 :
787 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
788 : pub enum GetTimelineError {
789 : #[error("Timeline is shutting down")]
790 : ShuttingDown,
791 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
792 : NotActive {
793 : tenant_id: TenantShardId,
794 : timeline_id: TimelineId,
795 : state: TimelineState,
796 : },
797 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
798 : NotFound {
799 : tenant_id: TenantShardId,
800 : timeline_id: TimelineId,
801 : },
802 : }
803 :
804 : #[derive(Debug, thiserror::Error)]
805 : pub enum LoadLocalTimelineError {
806 : #[error("FailedToLoad")]
807 : Load(#[source] anyhow::Error),
808 : #[error("FailedToResumeDeletion")]
809 : ResumeDeletion(#[source] anyhow::Error),
810 : }
811 :
812 : #[derive(thiserror::Error)]
813 : pub enum DeleteTimelineError {
814 : #[error("NotFound")]
815 : NotFound,
816 :
817 : #[error("HasChildren")]
818 : HasChildren(Vec<TimelineId>),
819 :
820 : #[error("Timeline deletion is already in progress")]
821 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
822 :
823 : #[error("Cancelled")]
824 : Cancelled,
825 :
826 : #[error(transparent)]
827 : Other(#[from] anyhow::Error),
828 : }
829 :
830 : impl Debug for DeleteTimelineError {
831 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
832 0 : match self {
833 0 : Self::NotFound => write!(f, "NotFound"),
834 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
835 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
836 0 : Self::Cancelled => f.debug_tuple("Cancelled").finish(),
837 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
838 : }
839 0 : }
840 : }
841 :
842 : #[derive(thiserror::Error)]
843 : pub enum TimelineArchivalError {
844 : #[error("NotFound")]
845 : NotFound,
846 :
847 : #[error("Timeout")]
848 : Timeout,
849 :
850 : #[error("Cancelled")]
851 : Cancelled,
852 :
853 : #[error("ancestor is archived: {}", .0)]
854 : HasArchivedParent(TimelineId),
855 :
856 : #[error("HasUnarchivedChildren")]
857 : HasUnarchivedChildren(Vec<TimelineId>),
858 :
859 : #[error("Timeline archival is already in progress")]
860 : AlreadyInProgress,
861 :
862 : #[error(transparent)]
863 : Other(anyhow::Error),
864 : }
865 :
866 : #[derive(thiserror::Error, Debug)]
867 : pub(crate) enum TenantManifestError {
868 : #[error("Remote storage error: {0}")]
869 : RemoteStorage(anyhow::Error),
870 :
871 : #[error("Cancelled")]
872 : Cancelled,
873 : }
874 :
875 : impl From<TenantManifestError> for TimelineArchivalError {
876 0 : fn from(e: TenantManifestError) -> Self {
877 0 : match e {
878 0 : TenantManifestError::RemoteStorage(e) => Self::Other(e),
879 0 : TenantManifestError::Cancelled => Self::Cancelled,
880 : }
881 0 : }
882 : }
883 :
884 : impl Debug for TimelineArchivalError {
885 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
886 0 : match self {
887 0 : Self::NotFound => write!(f, "NotFound"),
888 0 : Self::Timeout => write!(f, "Timeout"),
889 0 : Self::Cancelled => write!(f, "Cancelled"),
890 0 : Self::HasArchivedParent(p) => f.debug_tuple("HasArchivedParent").field(p).finish(),
891 0 : Self::HasUnarchivedChildren(c) => {
892 0 : f.debug_tuple("HasUnarchivedChildren").field(c).finish()
893 : }
894 0 : Self::AlreadyInProgress => f.debug_tuple("AlreadyInProgress").finish(),
895 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
896 : }
897 0 : }
898 : }
899 :
900 : pub enum SetStoppingError {
901 : AlreadyStopping(completion::Barrier),
902 : Broken,
903 : }
904 :
905 : impl Debug for SetStoppingError {
906 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
907 0 : match self {
908 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
909 0 : Self::Broken => write!(f, "Broken"),
910 : }
911 0 : }
912 : }
913 :
914 : #[derive(thiserror::Error, Debug)]
915 : pub(crate) enum FinalizeTimelineImportError {
916 : #[error("Import task not done yet")]
917 : ImportTaskStillRunning,
918 : #[error("Shutting down")]
919 : ShuttingDown,
920 : }
921 :
922 : /// Arguments to [`TenantShard::create_timeline`].
923 : ///
924 : /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
925 : /// is `None`, the result of the timeline create call is not deterministic.
926 : ///
927 : /// See [`CreateTimelineIdempotency`] for an idempotency key.
928 : #[derive(Debug)]
929 : pub(crate) enum CreateTimelineParams {
930 : Bootstrap(CreateTimelineParamsBootstrap),
931 : Branch(CreateTimelineParamsBranch),
932 : ImportPgdata(CreateTimelineParamsImportPgdata),
933 : }
934 :
935 : #[derive(Debug)]
936 : pub(crate) struct CreateTimelineParamsBootstrap {
937 : pub(crate) new_timeline_id: TimelineId,
938 : pub(crate) existing_initdb_timeline_id: Option<TimelineId>,
939 : pub(crate) pg_version: PgMajorVersion,
940 : }
941 :
942 : /// NB: See comment on [`CreateTimelineIdempotency::Branch`] for why there's no `pg_version` here.
943 : #[derive(Debug)]
944 : pub(crate) struct CreateTimelineParamsBranch {
945 : pub(crate) new_timeline_id: TimelineId,
946 : pub(crate) ancestor_timeline_id: TimelineId,
947 : pub(crate) ancestor_start_lsn: Option<Lsn>,
948 : }
949 :
950 : #[derive(Debug)]
951 : pub(crate) struct CreateTimelineParamsImportPgdata {
952 : pub(crate) new_timeline_id: TimelineId,
953 : pub(crate) location: import_pgdata::index_part_format::Location,
954 : pub(crate) idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
955 : }
956 :
957 : /// What is used to determine idempotency of a [`TenantShard::create_timeline`] call in [`TenantShard::start_creating_timeline`] in [`TenantShard::start_creating_timeline`].
958 : ///
959 : /// Each [`Timeline`] object holds [`Self`] as an immutable property in [`Timeline::create_idempotency`].
960 : ///
961 : /// We lower timeline creation requests to [`Self`], and then use [`PartialEq::eq`] to compare [`Timeline::create_idempotency`] with the request.
962 : /// If they are equal, we return a reference to the existing timeline, otherwise it's an idempotency conflict.
963 : ///
964 : /// There is special treatment for [`Self::FailWithConflict`] to always return an idempotency conflict.
965 : /// It would be nice to have more advanced derive macros to make that special treatment declarative.
966 : ///
967 : /// Notes:
968 : /// - Unlike [`CreateTimelineParams`], ancestor LSN is fixed, so, branching will be at a deterministic LSN.
969 : /// - We make some trade-offs though, e.g., [`CreateTimelineParamsBootstrap::existing_initdb_timeline_id`]
970 : /// is not considered for idempotency. We can improve on this over time if we deem it necessary.
971 : ///
972 : #[derive(Debug, Clone, PartialEq, Eq)]
973 : pub(crate) enum CreateTimelineIdempotency {
974 : /// NB: special treatment, see comment in [`Self`].
975 : FailWithConflict,
976 : Bootstrap {
977 : pg_version: PgMajorVersion,
978 : },
979 : /// NB: branches always have the same `pg_version` as their ancestor.
980 : /// While [`pageserver_api::models::TimelineCreateRequestMode::Branch::pg_version`]
981 : /// exists as a field, and is set by cplane, it has always been ignored by pageserver when
982 : /// determining the child branch pg_version.
983 : Branch {
984 : ancestor_timeline_id: TimelineId,
985 : ancestor_start_lsn: Lsn,
986 : },
987 : ImportPgdata(CreatingTimelineIdempotencyImportPgdata),
988 : }
989 :
990 : #[derive(Debug, Clone, PartialEq, Eq)]
991 : pub(crate) struct CreatingTimelineIdempotencyImportPgdata {
992 : idempotency_key: import_pgdata::index_part_format::IdempotencyKey,
993 : }
994 :
995 : /// What is returned by [`TenantShard::start_creating_timeline`].
996 : #[must_use]
997 : enum StartCreatingTimelineResult {
998 : CreateGuard(TimelineCreateGuard),
999 : Idempotent(Arc<Timeline>),
1000 : }
1001 :
1002 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1003 : enum TimelineInitAndSyncResult {
1004 : ReadyToActivate,
1005 : NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
1006 : }
1007 :
1008 : #[must_use]
1009 : struct TimelineInitAndSyncNeedsSpawnImportPgdata {
1010 : timeline: Arc<Timeline>,
1011 : import_pgdata: import_pgdata::index_part_format::Root,
1012 : guard: TimelineCreateGuard,
1013 : }
1014 :
1015 : /// What is returned by [`TenantShard::create_timeline`].
1016 : enum CreateTimelineResult {
1017 : Created(Arc<Timeline>),
1018 : Idempotent(Arc<Timeline>),
1019 : /// IMPORTANT: This [`Arc<Timeline>`] object is not in [`TenantShard::timelines`] when
1020 : /// we return this result, nor will this concrete object ever be added there.
1021 : /// Cf method comment on [`TenantShard::create_timeline_import_pgdata`].
1022 : ImportSpawned(Arc<Timeline>),
1023 : }
1024 :
1025 : impl CreateTimelineResult {
1026 0 : fn discriminant(&self) -> &'static str {
1027 0 : match self {
1028 0 : Self::Created(_) => "Created",
1029 0 : Self::Idempotent(_) => "Idempotent",
1030 0 : Self::ImportSpawned(_) => "ImportSpawned",
1031 : }
1032 0 : }
1033 0 : fn timeline(&self) -> &Arc<Timeline> {
1034 0 : match self {
1035 0 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
1036 : }
1037 0 : }
1038 : /// Unit test timelines aren't activated, test has to do it if it needs to.
1039 : #[cfg(test)]
1040 118 : fn into_timeline_for_test(self) -> Arc<Timeline> {
1041 118 : match self {
1042 118 : Self::Created(t) | Self::Idempotent(t) | Self::ImportSpawned(t) => t,
1043 : }
1044 118 : }
1045 : }
1046 :
1047 : #[derive(thiserror::Error, Debug)]
1048 : pub enum CreateTimelineError {
1049 : #[error("creation of timeline with the given ID is in progress")]
1050 : AlreadyCreating,
1051 : #[error("timeline already exists with different parameters")]
1052 : Conflict,
1053 : #[error(transparent)]
1054 : AncestorLsn(anyhow::Error),
1055 : #[error("ancestor timeline is not active")]
1056 : AncestorNotActive,
1057 : #[error("ancestor timeline is archived")]
1058 : AncestorArchived,
1059 : #[error("tenant shutting down")]
1060 : ShuttingDown,
1061 : #[error(transparent)]
1062 : Other(#[from] anyhow::Error),
1063 : }
1064 :
1065 : #[derive(thiserror::Error, Debug)]
1066 : pub enum InitdbError {
1067 : #[error("Operation was cancelled")]
1068 : Cancelled,
1069 : #[error(transparent)]
1070 : Other(anyhow::Error),
1071 : #[error(transparent)]
1072 : Inner(postgres_initdb::Error),
1073 : }
1074 :
1075 : enum CreateTimelineCause {
1076 : Load,
1077 : Delete,
1078 : }
1079 :
1080 : #[allow(clippy::large_enum_variant, reason = "TODO")]
1081 : enum LoadTimelineCause {
1082 : Attach,
1083 : Unoffload,
1084 : }
1085 :
1086 : #[derive(thiserror::Error, Debug)]
1087 : pub(crate) enum GcError {
1088 : // The tenant is shutting down
1089 : #[error("tenant shutting down")]
1090 : TenantCancelled,
1091 :
1092 : // The tenant is shutting down
1093 : #[error("timeline shutting down")]
1094 : TimelineCancelled,
1095 :
1096 : // The tenant is in a state inelegible to run GC
1097 : #[error("not active")]
1098 : NotActive,
1099 :
1100 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
1101 : #[error("not active")]
1102 : BadLsn { why: String },
1103 :
1104 : // A remote storage error while scheduling updates after compaction
1105 : #[error(transparent)]
1106 : Remote(anyhow::Error),
1107 :
1108 : // An error reading while calculating GC cutoffs
1109 : #[error(transparent)]
1110 : GcCutoffs(PageReconstructError),
1111 :
1112 : // If GC was invoked for a particular timeline, this error means it didn't exist
1113 : #[error("timeline not found")]
1114 : TimelineNotFound,
1115 : }
1116 :
1117 : impl From<PageReconstructError> for GcError {
1118 0 : fn from(value: PageReconstructError) -> Self {
1119 0 : match value {
1120 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
1121 0 : other => Self::GcCutoffs(other),
1122 : }
1123 0 : }
1124 : }
1125 :
1126 : impl From<NotInitialized> for GcError {
1127 0 : fn from(value: NotInitialized) -> Self {
1128 0 : match value {
1129 0 : NotInitialized::Uninitialized => GcError::Remote(value.into()),
1130 0 : NotInitialized::Stopped | NotInitialized::ShuttingDown => GcError::TimelineCancelled,
1131 : }
1132 0 : }
1133 : }
1134 :
1135 : impl From<timeline::layer_manager::Shutdown> for GcError {
1136 0 : fn from(_: timeline::layer_manager::Shutdown) -> Self {
1137 0 : GcError::TimelineCancelled
1138 0 : }
1139 : }
1140 :
1141 : #[derive(thiserror::Error, Debug)]
1142 : pub(crate) enum LoadConfigError {
1143 : #[error("TOML deserialization error: '{0}'")]
1144 : DeserializeToml(#[from] toml_edit::de::Error),
1145 :
1146 : #[error("Config not found at {0}")]
1147 : NotFound(Utf8PathBuf),
1148 : }
1149 :
1150 : impl TenantShard {
1151 : /// Yet another helper for timeline initialization.
1152 : ///
1153 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
1154 : /// - Scans the local timeline directory for layer files and builds the layer map
1155 : /// - Downloads remote index file and adds remote files to the layer map
1156 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
1157 : ///
1158 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
1159 : /// it is marked as Active.
1160 : #[allow(clippy::too_many_arguments)]
1161 3 : async fn timeline_init_and_sync(
1162 3 : self: &Arc<Self>,
1163 3 : timeline_id: TimelineId,
1164 3 : resources: TimelineResources,
1165 3 : index_part: IndexPart,
1166 3 : metadata: TimelineMetadata,
1167 3 : previous_heatmap: Option<PreviousHeatmap>,
1168 3 : ancestor: Option<Arc<Timeline>>,
1169 3 : cause: LoadTimelineCause,
1170 3 : ctx: &RequestContext,
1171 3 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
1172 3 : let tenant_id = self.tenant_shard_id;
1173 :
1174 3 : let import_pgdata = index_part.import_pgdata.clone();
1175 3 : let idempotency = match &import_pgdata {
1176 0 : Some(import_pgdata) => {
1177 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
1178 0 : idempotency_key: import_pgdata.idempotency_key().clone(),
1179 0 : })
1180 : }
1181 : None => {
1182 3 : if metadata.ancestor_timeline().is_none() {
1183 2 : CreateTimelineIdempotency::Bootstrap {
1184 2 : pg_version: metadata.pg_version(),
1185 2 : }
1186 : } else {
1187 1 : CreateTimelineIdempotency::Branch {
1188 1 : ancestor_timeline_id: metadata.ancestor_timeline().unwrap(),
1189 1 : ancestor_start_lsn: metadata.ancestor_lsn(),
1190 1 : }
1191 : }
1192 : }
1193 : };
1194 :
1195 3 : let (timeline, _timeline_ctx) = self.create_timeline_struct(
1196 3 : timeline_id,
1197 3 : &metadata,
1198 3 : previous_heatmap,
1199 3 : ancestor.clone(),
1200 3 : resources,
1201 3 : CreateTimelineCause::Load,
1202 3 : idempotency.clone(),
1203 3 : index_part.gc_compaction.clone(),
1204 3 : index_part.rel_size_migration.clone(),
1205 3 : ctx,
1206 3 : )?;
1207 3 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
1208 :
1209 3 : if !disk_consistent_lsn.is_valid() {
1210 : // As opposed to normal timelines which get initialised with a disk consitent LSN
1211 : // via initdb, imported timelines start from 0. If the import task stops before
1212 : // it advances disk consitent LSN, allow it to resume.
1213 0 : let in_progress_import = import_pgdata
1214 0 : .as_ref()
1215 0 : .map(|import| !import.is_done())
1216 0 : .unwrap_or(false);
1217 0 : if !in_progress_import {
1218 0 : anyhow::bail!("Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn");
1219 0 : }
1220 3 : }
1221 :
1222 3 : assert_eq!(
1223 : disk_consistent_lsn,
1224 3 : metadata.disk_consistent_lsn(),
1225 0 : "these are used interchangeably"
1226 : );
1227 :
1228 3 : timeline.remote_client.init_upload_queue(&index_part)?;
1229 :
1230 3 : timeline
1231 3 : .load_layer_map(disk_consistent_lsn, index_part)
1232 3 : .await
1233 3 : .with_context(|| {
1234 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
1235 0 : })?;
1236 :
1237 : // When unarchiving, we've mostly likely lost the heatmap generated prior
1238 : // to the archival operation. To allow warming this timeline up, generate
1239 : // a previous heatmap which contains all visible layers in the layer map.
1240 : // This previous heatmap will be used whenever a fresh heatmap is generated
1241 : // for the timeline.
1242 3 : if self.conf.generate_unarchival_heatmap && matches!(cause, LoadTimelineCause::Unoffload) {
1243 0 : let mut tline_ending_at = Some((&timeline, timeline.get_last_record_lsn()));
1244 0 : while let Some((tline, end_lsn)) = tline_ending_at {
1245 0 : let unarchival_heatmap = tline.generate_unarchival_heatmap(end_lsn).await;
1246 : // Another unearchived timeline might have generated a heatmap for this ancestor.
1247 : // If the current branch point greater than the previous one use the the heatmap
1248 : // we just generated - it should include more layers.
1249 0 : if !tline.should_keep_previous_heatmap(end_lsn) {
1250 0 : tline
1251 0 : .previous_heatmap
1252 0 : .store(Some(Arc::new(unarchival_heatmap)));
1253 0 : } else {
1254 0 : tracing::info!("Previous heatmap preferred. Dropping unarchival heatmap.")
1255 : }
1256 :
1257 0 : match tline.ancestor_timeline() {
1258 0 : Some(ancestor) => {
1259 0 : if ancestor.update_layer_visibility().await.is_err() {
1260 : // Ancestor timeline is shutting down.
1261 0 : break;
1262 0 : }
1263 :
1264 0 : tline_ending_at = Some((ancestor, tline.get_ancestor_lsn()));
1265 : }
1266 0 : None => {
1267 0 : tline_ending_at = None;
1268 0 : }
1269 : }
1270 : }
1271 3 : }
1272 :
1273 0 : match import_pgdata {
1274 0 : Some(import_pgdata) if !import_pgdata.is_done() => {
1275 0 : let mut guard = self.timelines_creating.lock().unwrap();
1276 0 : if !guard.insert(timeline_id) {
1277 : // We should never try and load the same timeline twice during startup
1278 0 : unreachable!("Timeline {tenant_id}/{timeline_id} is already being created")
1279 0 : }
1280 0 : let timeline_create_guard = TimelineCreateGuard {
1281 0 : _tenant_gate_guard: self.gate.enter()?,
1282 0 : owning_tenant: self.clone(),
1283 0 : timeline_id,
1284 0 : idempotency,
1285 : // The users of this specific return value don't need the timline_path in there.
1286 0 : timeline_path: timeline
1287 0 : .conf
1288 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id),
1289 : };
1290 0 : Ok(TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1291 0 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1292 0 : timeline,
1293 0 : import_pgdata,
1294 0 : guard: timeline_create_guard,
1295 0 : },
1296 0 : ))
1297 : }
1298 : Some(_) | None => {
1299 : {
1300 3 : let mut timelines_accessor = self.timelines.lock().unwrap();
1301 3 : match timelines_accessor.entry(timeline_id) {
1302 : // We should never try and load the same timeline twice during startup
1303 : Entry::Occupied(_) => {
1304 0 : unreachable!(
1305 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
1306 : );
1307 : }
1308 3 : Entry::Vacant(v) => {
1309 3 : v.insert(Arc::clone(&timeline));
1310 3 : timeline.maybe_spawn_flush_loop();
1311 3 : }
1312 : }
1313 : }
1314 :
1315 3 : if disk_consistent_lsn.is_valid() {
1316 : // Sanity check: a timeline should have some content.
1317 : // Exception: importing timelines might not yet have any
1318 3 : anyhow::ensure!(
1319 3 : ancestor.is_some()
1320 2 : || timeline
1321 2 : .layers
1322 2 : .read(LayerManagerLockHolder::LoadLayerMap)
1323 2 : .await
1324 2 : .layer_map()
1325 2 : .expect(
1326 2 : "currently loading, layer manager cannot be shutdown already"
1327 : )
1328 2 : .iter_historic_layers()
1329 2 : .next()
1330 2 : .is_some(),
1331 0 : "Timeline has no ancestor and no layer files"
1332 : );
1333 0 : }
1334 :
1335 3 : Ok(TimelineInitAndSyncResult::ReadyToActivate)
1336 : }
1337 : }
1338 3 : }
1339 :
1340 : /// Attach a tenant that's available in cloud storage.
1341 : ///
1342 : /// This returns quickly, after just creating the in-memory object
1343 : /// Tenant struct and launching a background task to download
1344 : /// the remote index files. On return, the tenant is most likely still in
1345 : /// Attaching state, and it will become Active once the background task
1346 : /// finishes. You can use wait_until_active() to wait for the task to
1347 : /// complete.
1348 : ///
1349 : #[allow(clippy::too_many_arguments)]
1350 0 : pub(crate) fn spawn(
1351 0 : conf: &'static PageServerConf,
1352 0 : tenant_shard_id: TenantShardId,
1353 0 : resources: TenantSharedResources,
1354 0 : attached_conf: AttachedTenantConf,
1355 0 : shard_identity: ShardIdentity,
1356 0 : init_order: Option<InitializationOrder>,
1357 0 : mode: SpawnMode,
1358 0 : ctx: &RequestContext,
1359 0 : ) -> Result<Arc<TenantShard>, GlobalShutDown> {
1360 0 : let wal_redo_manager =
1361 0 : WalRedoManager::new(PostgresRedoManager::new(conf, tenant_shard_id))?;
1362 :
1363 : let TenantSharedResources {
1364 0 : broker_client,
1365 0 : remote_storage,
1366 0 : deletion_queue_client,
1367 0 : l0_flush_global_state,
1368 0 : basebackup_cache,
1369 0 : feature_resolver,
1370 0 : } = resources;
1371 :
1372 0 : let attach_mode = attached_conf.location.attach_mode;
1373 0 : let generation = attached_conf.location.generation;
1374 :
1375 0 : let tenant = Arc::new(TenantShard::new(
1376 0 : TenantState::Attaching,
1377 0 : conf,
1378 0 : attached_conf,
1379 0 : shard_identity,
1380 0 : Some(wal_redo_manager),
1381 0 : tenant_shard_id,
1382 0 : remote_storage.clone(),
1383 0 : deletion_queue_client,
1384 0 : l0_flush_global_state,
1385 0 : basebackup_cache,
1386 0 : feature_resolver,
1387 : ));
1388 :
1389 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
1390 : // we shut down while attaching.
1391 0 : let attach_gate_guard = tenant
1392 0 : .gate
1393 0 : .enter()
1394 0 : .expect("We just created the TenantShard: nothing else can have shut it down yet");
1395 :
1396 : // Do all the hard work in the background
1397 0 : let tenant_clone = Arc::clone(&tenant);
1398 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
1399 0 : task_mgr::spawn(
1400 0 : &tokio::runtime::Handle::current(),
1401 0 : TaskKind::Attach,
1402 0 : tenant_shard_id,
1403 0 : None,
1404 0 : "attach tenant",
1405 0 : async move {
1406 :
1407 0 : info!(
1408 : ?attach_mode,
1409 0 : "Attaching tenant"
1410 : );
1411 :
1412 0 : let _gate_guard = attach_gate_guard;
1413 :
1414 : // Is this tenant being spawned as part of process startup?
1415 0 : let starting_up = init_order.is_some();
1416 0 : scopeguard::defer! {
1417 : if starting_up {
1418 : TENANT.startup_complete.inc();
1419 : }
1420 : }
1421 :
1422 0 : fn make_broken_or_stopping(t: &TenantShard, err: anyhow::Error) {
1423 0 : t.state.send_modify(|state| match state {
1424 : // TODO: the old code alluded to DeleteTenantFlow sometimes setting
1425 : // TenantState::Stopping before we get here, but this may be outdated.
1426 : // Let's find out with a testing assertion. If this doesn't fire, and the
1427 : // logs don't show this happening in production, remove the Stopping cases.
1428 0 : TenantState::Stopping{..} if cfg!(any(test, feature = "testing")) => {
1429 0 : panic!("unexpected TenantState::Stopping during attach")
1430 : }
1431 : // If the tenant is cancelled, assume the error was caused by cancellation.
1432 0 : TenantState::Attaching if t.cancel.is_cancelled() => {
1433 0 : info!("attach cancelled, setting tenant state to Stopping: {err}");
1434 : // NB: progress None tells `set_stopping` that attach has cancelled.
1435 0 : *state = TenantState::Stopping { progress: None };
1436 : }
1437 : // According to the old code, DeleteTenantFlow may already have set this to
1438 : // Stopping. Retain its progress.
1439 : // TODO: there is no DeleteTenantFlow. Is this still needed? See above.
1440 0 : TenantState::Stopping { progress } if t.cancel.is_cancelled() => {
1441 0 : assert!(progress.is_some(), "concurrent attach cancellation");
1442 0 : info!("attach cancelled, already Stopping: {err}");
1443 : }
1444 : // Mark the tenant as broken.
1445 : TenantState::Attaching | TenantState::Stopping { .. } => {
1446 0 : error!("attach failed, setting tenant state to Broken (was {state}): {err:?}");
1447 0 : *state = TenantState::broken_from_reason(err.to_string())
1448 : }
1449 : // The attach task owns the tenant state until activated.
1450 0 : state => panic!("invalid tenant state {state} during attach: {err:?}"),
1451 0 : });
1452 0 : }
1453 :
1454 : // TODO: should also be rejecting tenant conf changes that violate this check.
1455 0 : if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
1456 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
1457 0 : return Ok(());
1458 0 : }
1459 :
1460 0 : let mut init_order = init_order;
1461 : // take the completion because initial tenant loading will complete when all of
1462 : // these tasks complete.
1463 0 : let _completion = init_order
1464 0 : .as_mut()
1465 0 : .and_then(|x| x.initial_tenant_load.take());
1466 0 : let remote_load_completion = init_order
1467 0 : .as_mut()
1468 0 : .and_then(|x| x.initial_tenant_load_remote.take());
1469 :
1470 : enum AttachType<'a> {
1471 : /// We are attaching this tenant lazily in the background.
1472 : Warmup {
1473 : _permit: tokio::sync::SemaphorePermit<'a>,
1474 : during_startup: bool
1475 : },
1476 : /// We are attaching this tenant as soon as we can, because for example an
1477 : /// endpoint tried to access it.
1478 : OnDemand,
1479 : /// During normal operations after startup, we are attaching a tenant, and
1480 : /// eager attach was requested.
1481 : Normal,
1482 : }
1483 :
1484 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
1485 : // Before doing any I/O, wait for at least one of:
1486 : // - A client attempting to access to this tenant (on-demand loading)
1487 : // - A permit becoming available in the warmup semaphore (background warmup)
1488 :
1489 0 : tokio::select!(
1490 0 : permit = tenant_clone.activate_now_sem.acquire() => {
1491 0 : let _ = permit.expect("activate_now_sem is never closed");
1492 0 : tracing::info!("Activating tenant (on-demand)");
1493 0 : AttachType::OnDemand
1494 : },
1495 0 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
1496 0 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
1497 0 : tracing::info!("Activating tenant (warmup)");
1498 0 : AttachType::Warmup {
1499 0 : _permit,
1500 0 : during_startup: init_order.is_some()
1501 0 : }
1502 : }
1503 0 : _ = tenant_clone.cancel.cancelled() => {
1504 : // This is safe, but should be pretty rare: it is interesting if a tenant
1505 : // stayed in Activating for such a long time that shutdown found it in
1506 : // that state.
1507 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
1508 : // Set the tenant to Stopping to signal `set_stopping` that we're done.
1509 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"));
1510 0 : return Ok(());
1511 : },
1512 : )
1513 : } else {
1514 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
1515 : // concurrent_tenant_warmup queue
1516 0 : AttachType::Normal
1517 : };
1518 :
1519 0 : let preload = match &mode {
1520 : SpawnMode::Eager | SpawnMode::Lazy => {
1521 0 : let _preload_timer = TENANT.preload.start_timer();
1522 0 : let res = tenant_clone
1523 0 : .preload(&remote_storage, task_mgr::shutdown_token())
1524 0 : .await;
1525 0 : match res {
1526 0 : Ok(p) => Some(p),
1527 0 : Err(e) => {
1528 0 : make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
1529 0 : return Ok(());
1530 : }
1531 : }
1532 : }
1533 :
1534 : };
1535 :
1536 : // Remote preload is complete.
1537 0 : drop(remote_load_completion);
1538 :
1539 :
1540 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
1541 0 : let attach_start = std::time::Instant::now();
1542 0 : let attached = {
1543 0 : let _attach_timer = Some(TENANT.attach.start_timer());
1544 0 : tenant_clone.attach(preload, &ctx).await
1545 : };
1546 0 : let attach_duration = attach_start.elapsed();
1547 0 : _ = tenant_clone.attach_wal_lag_cooldown.set(WalLagCooldown::new(attach_start, attach_duration));
1548 :
1549 0 : match attached {
1550 : Ok(()) => {
1551 0 : info!("attach finished, activating");
1552 0 : tenant_clone.activate(broker_client, None, &ctx);
1553 : }
1554 0 : Err(e) => make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e)),
1555 : }
1556 :
1557 : // If we are doing an opportunistic warmup attachment at startup, initialize
1558 : // logical size at the same time. This is better than starting a bunch of idle tenants
1559 : // with cold caches and then coming back later to initialize their logical sizes.
1560 : //
1561 : // It also prevents the warmup proccess competing with the concurrency limit on
1562 : // logical size calculations: if logical size calculation semaphore is saturated,
1563 : // then warmup will wait for that before proceeding to the next tenant.
1564 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
1565 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
1566 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
1567 0 : while futs.next().await.is_some() {}
1568 0 : tracing::info!("Warm-up complete");
1569 0 : }
1570 :
1571 0 : Ok(())
1572 0 : }
1573 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
1574 : );
1575 0 : Ok(tenant)
1576 0 : }
1577 :
1578 : #[instrument(skip_all)]
1579 : pub(crate) async fn preload(
1580 : self: &Arc<Self>,
1581 : remote_storage: &GenericRemoteStorage,
1582 : cancel: CancellationToken,
1583 : ) -> anyhow::Result<TenantPreload> {
1584 : span::debug_assert_current_span_has_tenant_id();
1585 : // Get list of remote timelines
1586 : // download index files for every tenant timeline
1587 : info!("listing remote timelines");
1588 : let (mut remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
1589 : remote_storage,
1590 : self.tenant_shard_id,
1591 : cancel.clone(),
1592 : )
1593 : .await?;
1594 :
1595 : let tenant_manifest = match download_tenant_manifest(
1596 : remote_storage,
1597 : &self.tenant_shard_id,
1598 : self.generation,
1599 : &cancel,
1600 : )
1601 : .await
1602 : {
1603 : Ok((tenant_manifest, _, _)) => Some(tenant_manifest),
1604 : Err(DownloadError::NotFound) => None,
1605 : Err(err) => return Err(err.into()),
1606 : };
1607 :
1608 : info!(
1609 : "found {} timelines ({} offloaded timelines)",
1610 : remote_timeline_ids.len(),
1611 : tenant_manifest
1612 : .as_ref()
1613 3 : .map(|m| m.offloaded_timelines.len())
1614 : .unwrap_or(0)
1615 : );
1616 :
1617 : for k in other_keys {
1618 : warn!("Unexpected non timeline key {k}");
1619 : }
1620 :
1621 : // Avoid downloading IndexPart of offloaded timelines.
1622 : let mut offloaded_with_prefix = HashSet::new();
1623 : if let Some(tenant_manifest) = &tenant_manifest {
1624 : for offloaded in tenant_manifest.offloaded_timelines.iter() {
1625 : if remote_timeline_ids.remove(&offloaded.timeline_id) {
1626 : offloaded_with_prefix.insert(offloaded.timeline_id);
1627 : } else {
1628 : // We'll take care later of timelines in the manifest without a prefix
1629 : }
1630 : }
1631 : }
1632 :
1633 : // TODO(vlad): Could go to S3 if the secondary is freezing cold and hasn't even
1634 : // pulled the first heatmap. Not entirely necessary since the storage controller
1635 : // will kick the secondary in any case and cause a download.
1636 : let maybe_heatmap_at = self.read_on_disk_heatmap().await;
1637 :
1638 : let timelines = self
1639 : .load_timelines_metadata(
1640 : remote_timeline_ids,
1641 : remote_storage,
1642 : maybe_heatmap_at,
1643 : cancel,
1644 : )
1645 : .await?;
1646 :
1647 : Ok(TenantPreload {
1648 : tenant_manifest,
1649 : timelines: timelines
1650 : .into_iter()
1651 3 : .map(|(id, tl)| (id, Some(tl)))
1652 0 : .chain(offloaded_with_prefix.into_iter().map(|id| (id, None)))
1653 : .collect(),
1654 : })
1655 : }
1656 :
1657 118 : async fn read_on_disk_heatmap(&self) -> Option<(HeatMapTenant, std::time::Instant)> {
1658 118 : if !self.conf.load_previous_heatmap {
1659 0 : return None;
1660 118 : }
1661 :
1662 118 : let on_disk_heatmap_path = self.conf.tenant_heatmap_path(&self.tenant_shard_id);
1663 118 : match tokio::fs::read_to_string(on_disk_heatmap_path).await {
1664 0 : Ok(heatmap) => match serde_json::from_str::<HeatMapTenant>(&heatmap) {
1665 0 : Ok(heatmap) => Some((heatmap, std::time::Instant::now())),
1666 0 : Err(err) => {
1667 0 : error!("Failed to deserialize old heatmap: {err}");
1668 0 : None
1669 : }
1670 : },
1671 118 : Err(err) => match err.kind() {
1672 118 : std::io::ErrorKind::NotFound => None,
1673 : _ => {
1674 0 : error!("Unexpected IO error reading old heatmap: {err}");
1675 0 : None
1676 : }
1677 : },
1678 : }
1679 118 : }
1680 :
1681 : ///
1682 : /// Background task that downloads all data for a tenant and brings it to Active state.
1683 : ///
1684 : /// No background tasks are started as part of this routine.
1685 : ///
1686 118 : async fn attach(
1687 118 : self: &Arc<TenantShard>,
1688 118 : preload: Option<TenantPreload>,
1689 118 : ctx: &RequestContext,
1690 118 : ) -> anyhow::Result<()> {
1691 118 : span::debug_assert_current_span_has_tenant_id();
1692 :
1693 118 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
1694 :
1695 118 : let Some(preload) = preload else {
1696 0 : anyhow::bail!(
1697 0 : "local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624"
1698 : );
1699 : };
1700 :
1701 118 : let mut offloaded_timeline_ids = HashSet::new();
1702 118 : let mut offloaded_timelines_list = Vec::new();
1703 118 : if let Some(tenant_manifest) = &preload.tenant_manifest {
1704 3 : for timeline_manifest in tenant_manifest.offloaded_timelines.iter() {
1705 0 : let timeline_id = timeline_manifest.timeline_id;
1706 0 : let offloaded_timeline =
1707 0 : OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
1708 0 : offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
1709 0 : offloaded_timeline_ids.insert(timeline_id);
1710 0 : }
1711 115 : }
1712 : // Complete deletions for offloaded timeline id's from manifest.
1713 : // The manifest will be uploaded later in this function.
1714 118 : offloaded_timelines_list
1715 118 : .retain(|(offloaded_id, offloaded)| {
1716 : // Existence of a timeline is finally determined by the existence of an index-part.json in remote storage.
1717 : // If there is dangling references in another location, they need to be cleaned up.
1718 0 : let delete = !preload.timelines.contains_key(offloaded_id);
1719 0 : if delete {
1720 0 : tracing::info!("Removing offloaded timeline {offloaded_id} from manifest as no remote prefix was found");
1721 0 : offloaded.defuse_for_tenant_drop();
1722 0 : }
1723 0 : !delete
1724 0 : });
1725 :
1726 118 : let mut timelines_to_resume_deletions = vec![];
1727 :
1728 118 : let mut remote_index_and_client = HashMap::new();
1729 118 : let mut timeline_ancestors = HashMap::new();
1730 118 : let mut existent_timelines = HashSet::new();
1731 121 : for (timeline_id, preload) in preload.timelines {
1732 3 : let Some(preload) = preload else { continue };
1733 : // This is an invariant of the `preload` function's API
1734 3 : assert!(!offloaded_timeline_ids.contains(&timeline_id));
1735 3 : let index_part = match preload.index_part {
1736 3 : Ok(i) => {
1737 3 : debug!("remote index part exists for timeline {timeline_id}");
1738 : // We found index_part on the remote, this is the standard case.
1739 3 : existent_timelines.insert(timeline_id);
1740 3 : i
1741 : }
1742 : Err(DownloadError::NotFound) => {
1743 : // There is no index_part on the remote. We only get here
1744 : // if there is some prefix for the timeline in the remote storage.
1745 : // This can e.g. be the initdb.tar.zst archive, maybe a
1746 : // remnant from a prior incomplete creation or deletion attempt.
1747 : // Delete the local directory as the deciding criterion for a
1748 : // timeline's existence is presence of index_part.
1749 0 : info!(%timeline_id, "index_part not found on remote");
1750 0 : continue;
1751 : }
1752 0 : Err(DownloadError::Fatal(why)) => {
1753 : // If, while loading one remote timeline, we saw an indication that our generation
1754 : // number is likely invalid, then we should not load the whole tenant.
1755 0 : error!(%timeline_id, "Fatal error loading timeline: {why}");
1756 0 : anyhow::bail!(why.to_string());
1757 : }
1758 0 : Err(e) => {
1759 : // Some (possibly ephemeral) error happened during index_part download.
1760 : // Pretend the timeline exists to not delete the timeline directory,
1761 : // as it might be a temporary issue and we don't want to re-download
1762 : // everything after it resolves.
1763 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
1764 :
1765 0 : existent_timelines.insert(timeline_id);
1766 0 : continue;
1767 : }
1768 : };
1769 3 : match index_part {
1770 3 : MaybeDeletedIndexPart::IndexPart(index_part) => {
1771 3 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
1772 3 : remote_index_and_client.insert(
1773 3 : timeline_id,
1774 3 : (index_part, preload.client, preload.previous_heatmap),
1775 3 : );
1776 3 : }
1777 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
1778 0 : info!(
1779 0 : "timeline {} is deleted, picking to resume deletion",
1780 : timeline_id
1781 : );
1782 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
1783 : }
1784 : }
1785 : }
1786 :
1787 118 : let mut gc_blocks = HashMap::new();
1788 :
1789 : // For every timeline, download the metadata file, scan the local directory,
1790 : // and build a layer map that contains an entry for each remote and local
1791 : // layer file.
1792 118 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
1793 121 : for (timeline_id, remote_metadata) in sorted_timelines {
1794 3 : let (index_part, remote_client, previous_heatmap) = remote_index_and_client
1795 3 : .remove(&timeline_id)
1796 3 : .expect("just put it in above");
1797 :
1798 3 : if let Some(blocking) = index_part.gc_blocking.as_ref() {
1799 : // could just filter these away, but it helps while testing
1800 0 : anyhow::ensure!(
1801 0 : !blocking.reasons.is_empty(),
1802 0 : "index_part for {timeline_id} is malformed: it should not have gc blocking with zero reasons"
1803 : );
1804 0 : let prev = gc_blocks.insert(timeline_id, blocking.reasons);
1805 0 : assert!(prev.is_none());
1806 3 : }
1807 :
1808 : // TODO again handle early failure
1809 3 : let effect = self
1810 3 : .load_remote_timeline(
1811 3 : timeline_id,
1812 3 : index_part,
1813 3 : remote_metadata,
1814 3 : previous_heatmap,
1815 3 : self.get_timeline_resources_for(remote_client),
1816 3 : LoadTimelineCause::Attach,
1817 3 : ctx,
1818 3 : )
1819 3 : .await
1820 3 : .with_context(|| {
1821 0 : format!(
1822 0 : "failed to load remote timeline {} for tenant {}",
1823 0 : timeline_id, self.tenant_shard_id
1824 : )
1825 0 : })?;
1826 :
1827 3 : match effect {
1828 3 : TimelineInitAndSyncResult::ReadyToActivate => {
1829 3 : // activation happens later, on Tenant::activate
1830 3 : }
1831 : TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
1832 : TimelineInitAndSyncNeedsSpawnImportPgdata {
1833 0 : timeline,
1834 0 : import_pgdata,
1835 0 : guard,
1836 : },
1837 : ) => {
1838 0 : let timeline_id = timeline.timeline_id;
1839 0 : let import_task_gate = Gate::default();
1840 0 : let import_task_guard = import_task_gate.enter().unwrap();
1841 0 : let import_task_handle =
1842 0 : tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
1843 0 : timeline.clone(),
1844 0 : import_pgdata,
1845 0 : guard,
1846 0 : import_task_guard,
1847 0 : ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
1848 : ));
1849 :
1850 0 : let prev = self.timelines_importing.lock().unwrap().insert(
1851 0 : timeline_id,
1852 0 : Arc::new(ImportingTimeline {
1853 0 : timeline: timeline.clone(),
1854 0 : import_task_handle,
1855 0 : import_task_gate,
1856 0 : delete_progress: TimelineDeleteProgress::default(),
1857 0 : }),
1858 0 : );
1859 :
1860 0 : assert!(prev.is_none());
1861 : }
1862 : }
1863 : }
1864 :
1865 : // At this point we've initialized all timelines and are tracking them.
1866 : // Now compute the layer visibility for all (not offloaded) timelines.
1867 118 : let compute_visiblity_for = {
1868 118 : let timelines_accessor = self.timelines.lock().unwrap();
1869 118 : let mut timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
1870 :
1871 118 : timelines_offloaded_accessor.extend(offloaded_timelines_list.into_iter());
1872 :
1873 : // Before activation, populate each Timeline's GcInfo with information about its children
1874 118 : self.initialize_gc_info(&timelines_accessor, &timelines_offloaded_accessor, None);
1875 :
1876 118 : timelines_accessor.values().cloned().collect::<Vec<_>>()
1877 : };
1878 :
1879 121 : for tl in compute_visiblity_for {
1880 3 : tl.update_layer_visibility().await.with_context(|| {
1881 0 : format!(
1882 0 : "failed initial timeline visibility computation {} for tenant {}",
1883 0 : tl.timeline_id, self.tenant_shard_id
1884 : )
1885 0 : })?;
1886 : }
1887 :
1888 : // Walk through deleted timelines, resume deletion
1889 118 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1890 0 : remote_timeline_client
1891 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1892 0 : .context("init queue stopped")
1893 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1894 :
1895 0 : DeleteTimelineFlow::resume_deletion(
1896 0 : Arc::clone(self),
1897 0 : timeline_id,
1898 0 : &index_part.metadata,
1899 0 : remote_timeline_client,
1900 0 : ctx,
1901 : )
1902 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1903 0 : .await
1904 0 : .context("resume_deletion")
1905 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1906 : }
1907 :
1908 : // Stash the preloaded tenant manifest, and upload a new manifest if changed.
1909 : //
1910 : // NB: this must happen after the tenant is fully populated above. In particular the
1911 : // offloaded timelines, which are included in the manifest.
1912 : {
1913 118 : let mut guard = self.remote_tenant_manifest.lock().await;
1914 118 : assert!(guard.is_none(), "tenant manifest set before preload"); // first populated here
1915 118 : *guard = preload.tenant_manifest;
1916 : }
1917 118 : self.maybe_upload_tenant_manifest().await?;
1918 :
1919 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1920 : // IndexPart is the source of truth.
1921 118 : self.clean_up_timelines(&existent_timelines)?;
1922 :
1923 118 : self.gc_block.set_scanned(gc_blocks);
1924 :
1925 118 : fail::fail_point!("attach-before-activate", |_| {
1926 0 : anyhow::bail!("attach-before-activate");
1927 0 : });
1928 118 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1929 :
1930 118 : info!("Done");
1931 :
1932 118 : Ok(())
1933 118 : }
1934 :
1935 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1936 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1937 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1938 118 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1939 118 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1940 :
1941 118 : let entries = match timelines_dir.read_dir_utf8() {
1942 118 : Ok(d) => d,
1943 0 : Err(e) => {
1944 0 : if e.kind() == std::io::ErrorKind::NotFound {
1945 0 : return Ok(());
1946 : } else {
1947 0 : return Err(e).context("list timelines directory for tenant");
1948 : }
1949 : }
1950 : };
1951 :
1952 122 : for entry in entries {
1953 4 : let entry = entry.context("read timeline dir entry")?;
1954 4 : let entry_path = entry.path();
1955 :
1956 4 : let purge = if crate::is_temporary(entry_path) {
1957 0 : true
1958 : } else {
1959 4 : match TimelineId::try_from(entry_path.file_name()) {
1960 4 : Ok(i) => {
1961 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1962 4 : !existent_timelines.contains(&i)
1963 : }
1964 0 : Err(e) => {
1965 0 : tracing::warn!(
1966 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1967 : );
1968 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1969 0 : false
1970 : }
1971 : }
1972 : };
1973 :
1974 4 : if purge {
1975 1 : tracing::info!("Purging stale timeline dentry {entry_path}");
1976 1 : if let Err(e) = match entry.file_type() {
1977 1 : Ok(t) => if t.is_dir() {
1978 1 : std::fs::remove_dir_all(entry_path)
1979 : } else {
1980 0 : std::fs::remove_file(entry_path)
1981 : }
1982 1 : .or_else(fs_ext::ignore_not_found),
1983 0 : Err(e) => Err(e),
1984 : } {
1985 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1986 1 : }
1987 3 : }
1988 : }
1989 :
1990 118 : Ok(())
1991 118 : }
1992 :
1993 : /// Get sum of all remote timelines sizes
1994 : ///
1995 : /// This function relies on the index_part instead of listing the remote storage
1996 0 : pub fn remote_size(&self) -> u64 {
1997 0 : let mut size = 0;
1998 :
1999 0 : for timeline in self.list_timelines() {
2000 0 : size += timeline.remote_client.get_remote_physical_size();
2001 0 : }
2002 :
2003 0 : size
2004 0 : }
2005 :
2006 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
2007 : #[allow(clippy::too_many_arguments)]
2008 : async fn load_remote_timeline(
2009 : self: &Arc<Self>,
2010 : timeline_id: TimelineId,
2011 : index_part: IndexPart,
2012 : remote_metadata: TimelineMetadata,
2013 : previous_heatmap: Option<PreviousHeatmap>,
2014 : resources: TimelineResources,
2015 : cause: LoadTimelineCause,
2016 : ctx: &RequestContext,
2017 : ) -> anyhow::Result<TimelineInitAndSyncResult> {
2018 : span::debug_assert_current_span_has_tenant_id();
2019 :
2020 : info!("downloading index file for timeline {}", timeline_id);
2021 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
2022 : .await
2023 : .context("Failed to create new timeline directory")?;
2024 :
2025 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
2026 : let timelines = self.timelines.lock().unwrap();
2027 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
2028 0 : || {
2029 0 : anyhow::anyhow!(
2030 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
2031 : )
2032 0 : },
2033 : )?))
2034 : } else {
2035 : None
2036 : };
2037 :
2038 : self.timeline_init_and_sync(
2039 : timeline_id,
2040 : resources,
2041 : index_part,
2042 : remote_metadata,
2043 : previous_heatmap,
2044 : ancestor,
2045 : cause,
2046 : ctx,
2047 : )
2048 : .await
2049 : }
2050 :
2051 118 : async fn load_timelines_metadata(
2052 118 : self: &Arc<TenantShard>,
2053 118 : timeline_ids: HashSet<TimelineId>,
2054 118 : remote_storage: &GenericRemoteStorage,
2055 118 : heatmap: Option<(HeatMapTenant, std::time::Instant)>,
2056 118 : cancel: CancellationToken,
2057 118 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
2058 118 : let mut timeline_heatmaps = heatmap.map(|h| (h.0.into_timelines_index(), h.1));
2059 :
2060 118 : let mut part_downloads = JoinSet::new();
2061 121 : for timeline_id in timeline_ids {
2062 3 : let cancel_clone = cancel.clone();
2063 :
2064 3 : let previous_timeline_heatmap = timeline_heatmaps.as_mut().and_then(|hs| {
2065 0 : hs.0.remove(&timeline_id).map(|h| PreviousHeatmap::Active {
2066 0 : heatmap: h,
2067 0 : read_at: hs.1,
2068 0 : end_lsn: None,
2069 0 : })
2070 0 : });
2071 3 : part_downloads.spawn(
2072 3 : self.load_timeline_metadata(
2073 3 : timeline_id,
2074 3 : remote_storage.clone(),
2075 3 : previous_timeline_heatmap,
2076 3 : cancel_clone,
2077 : )
2078 3 : .instrument(info_span!("download_index_part", %timeline_id)),
2079 : );
2080 : }
2081 :
2082 118 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
2083 :
2084 : loop {
2085 121 : tokio::select!(
2086 121 : next = part_downloads.join_next() => {
2087 121 : match next {
2088 3 : Some(result) => {
2089 3 : let preload = result.context("join preload task")?;
2090 3 : timeline_preloads.insert(preload.timeline_id, preload);
2091 : },
2092 : None => {
2093 118 : break;
2094 : }
2095 : }
2096 : },
2097 121 : _ = cancel.cancelled() => {
2098 0 : anyhow::bail!("Cancelled while waiting for remote index download")
2099 : }
2100 : )
2101 : }
2102 :
2103 118 : Ok(timeline_preloads)
2104 118 : }
2105 :
2106 3 : fn build_timeline_client(
2107 3 : &self,
2108 3 : timeline_id: TimelineId,
2109 3 : remote_storage: GenericRemoteStorage,
2110 3 : ) -> RemoteTimelineClient {
2111 3 : RemoteTimelineClient::new(
2112 3 : remote_storage.clone(),
2113 3 : self.deletion_queue_client.clone(),
2114 3 : self.conf,
2115 3 : self.tenant_shard_id,
2116 3 : timeline_id,
2117 3 : self.generation,
2118 3 : &self.tenant_conf.load().location,
2119 : )
2120 3 : }
2121 :
2122 3 : fn load_timeline_metadata(
2123 3 : self: &Arc<TenantShard>,
2124 3 : timeline_id: TimelineId,
2125 3 : remote_storage: GenericRemoteStorage,
2126 3 : previous_heatmap: Option<PreviousHeatmap>,
2127 3 : cancel: CancellationToken,
2128 3 : ) -> impl Future<Output = TimelinePreload> + use<> {
2129 3 : let client = self.build_timeline_client(timeline_id, remote_storage);
2130 3 : async move {
2131 3 : debug_assert_current_span_has_tenant_and_timeline_id();
2132 3 : debug!("starting index part download");
2133 :
2134 3 : let index_part = client.download_index_file(&cancel).await;
2135 :
2136 3 : debug!("finished index part download");
2137 :
2138 3 : TimelinePreload {
2139 3 : client,
2140 3 : timeline_id,
2141 3 : index_part,
2142 3 : previous_heatmap,
2143 3 : }
2144 3 : }
2145 3 : }
2146 :
2147 0 : fn check_to_be_archived_has_no_unarchived_children(
2148 0 : timeline_id: TimelineId,
2149 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2150 0 : ) -> Result<(), TimelineArchivalError> {
2151 0 : let children: Vec<TimelineId> = timelines
2152 0 : .iter()
2153 0 : .filter_map(|(id, entry)| {
2154 0 : if entry.get_ancestor_timeline_id() != Some(timeline_id) {
2155 0 : return None;
2156 0 : }
2157 0 : if entry.is_archived() == Some(true) {
2158 0 : return None;
2159 0 : }
2160 0 : Some(*id)
2161 0 : })
2162 0 : .collect();
2163 :
2164 0 : if !children.is_empty() {
2165 0 : return Err(TimelineArchivalError::HasUnarchivedChildren(children));
2166 0 : }
2167 0 : Ok(())
2168 0 : }
2169 :
2170 0 : fn check_ancestor_of_to_be_unarchived_is_not_archived(
2171 0 : ancestor_timeline_id: TimelineId,
2172 0 : timelines: &std::sync::MutexGuard<'_, HashMap<TimelineId, Arc<Timeline>>>,
2173 0 : offloaded_timelines: &std::sync::MutexGuard<
2174 0 : '_,
2175 0 : HashMap<TimelineId, Arc<OffloadedTimeline>>,
2176 0 : >,
2177 0 : ) -> Result<(), TimelineArchivalError> {
2178 0 : let has_archived_parent =
2179 0 : if let Some(ancestor_timeline) = timelines.get(&ancestor_timeline_id) {
2180 0 : ancestor_timeline.is_archived() == Some(true)
2181 0 : } else if offloaded_timelines.contains_key(&ancestor_timeline_id) {
2182 0 : true
2183 : } else {
2184 0 : error!("ancestor timeline {ancestor_timeline_id} not found");
2185 0 : if cfg!(debug_assertions) {
2186 0 : panic!("ancestor timeline {ancestor_timeline_id} not found");
2187 0 : }
2188 0 : return Err(TimelineArchivalError::NotFound);
2189 : };
2190 0 : if has_archived_parent {
2191 0 : return Err(TimelineArchivalError::HasArchivedParent(
2192 0 : ancestor_timeline_id,
2193 0 : ));
2194 0 : }
2195 0 : Ok(())
2196 0 : }
2197 :
2198 0 : fn check_to_be_unarchived_timeline_has_no_archived_parent(
2199 0 : timeline: &Arc<Timeline>,
2200 0 : ) -> Result<(), TimelineArchivalError> {
2201 0 : if let Some(ancestor_timeline) = timeline.ancestor_timeline() {
2202 0 : if ancestor_timeline.is_archived() == Some(true) {
2203 0 : return Err(TimelineArchivalError::HasArchivedParent(
2204 0 : ancestor_timeline.timeline_id,
2205 0 : ));
2206 0 : }
2207 0 : }
2208 0 : Ok(())
2209 0 : }
2210 :
2211 : /// Loads the specified (offloaded) timeline from S3 and attaches it as a loaded timeline
2212 : ///
2213 : /// Counterpart to [`offload_timeline`].
2214 0 : async fn unoffload_timeline(
2215 0 : self: &Arc<Self>,
2216 0 : timeline_id: TimelineId,
2217 0 : broker_client: storage_broker::BrokerClientChannel,
2218 0 : ctx: RequestContext,
2219 0 : ) -> Result<Arc<Timeline>, TimelineArchivalError> {
2220 0 : info!("unoffloading timeline");
2221 :
2222 : // We activate the timeline below manually, so this must be called on an active tenant.
2223 : // We expect callers of this function to ensure this.
2224 0 : match self.current_state() {
2225 : TenantState::Activating { .. }
2226 : | TenantState::Attaching
2227 : | TenantState::Broken { .. } => {
2228 0 : panic!("Timeline expected to be active")
2229 : }
2230 0 : TenantState::Stopping { .. } => return Err(TimelineArchivalError::Cancelled),
2231 0 : TenantState::Active => {}
2232 : }
2233 0 : let cancel = self.cancel.clone();
2234 :
2235 : // Protect against concurrent attempts to use this TimelineId
2236 : // We don't care much about idempotency, as it's ensured a layer above.
2237 0 : let allow_offloaded = true;
2238 0 : let _create_guard = self
2239 0 : .create_timeline_create_guard(
2240 0 : timeline_id,
2241 0 : CreateTimelineIdempotency::FailWithConflict,
2242 0 : allow_offloaded,
2243 : )
2244 0 : .map_err(|err| match err {
2245 0 : TimelineExclusionError::AlreadyCreating => TimelineArchivalError::AlreadyInProgress,
2246 : TimelineExclusionError::AlreadyExists { .. } => {
2247 0 : TimelineArchivalError::Other(anyhow::anyhow!("Timeline already exists"))
2248 : }
2249 0 : TimelineExclusionError::Other(e) => TimelineArchivalError::Other(e),
2250 0 : TimelineExclusionError::ShuttingDown => TimelineArchivalError::Cancelled,
2251 0 : })?;
2252 :
2253 0 : let timeline_preload = self
2254 0 : .load_timeline_metadata(
2255 0 : timeline_id,
2256 0 : self.remote_storage.clone(),
2257 0 : None,
2258 0 : cancel.clone(),
2259 0 : )
2260 0 : .await;
2261 :
2262 0 : let index_part = match timeline_preload.index_part {
2263 0 : Ok(index_part) => {
2264 0 : debug!("remote index part exists for timeline {timeline_id}");
2265 0 : index_part
2266 : }
2267 : Err(DownloadError::NotFound) => {
2268 0 : error!(%timeline_id, "index_part not found on remote");
2269 0 : return Err(TimelineArchivalError::NotFound);
2270 : }
2271 0 : Err(DownloadError::Cancelled) => return Err(TimelineArchivalError::Cancelled),
2272 0 : Err(e) => {
2273 : // Some (possibly ephemeral) error happened during index_part download.
2274 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
2275 0 : return Err(TimelineArchivalError::Other(
2276 0 : anyhow::Error::new(e).context("downloading index_part from remote storage"),
2277 0 : ));
2278 : }
2279 : };
2280 0 : let index_part = match index_part {
2281 0 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2282 0 : MaybeDeletedIndexPart::Deleted(_index_part) => {
2283 0 : info!("timeline is deleted according to index_part.json");
2284 0 : return Err(TimelineArchivalError::NotFound);
2285 : }
2286 : };
2287 0 : let remote_metadata = index_part.metadata.clone();
2288 0 : let timeline_resources = self.build_timeline_resources(timeline_id);
2289 0 : self.load_remote_timeline(
2290 0 : timeline_id,
2291 0 : index_part,
2292 0 : remote_metadata,
2293 0 : None,
2294 0 : timeline_resources,
2295 0 : LoadTimelineCause::Unoffload,
2296 0 : &ctx,
2297 0 : )
2298 0 : .await
2299 0 : .with_context(|| {
2300 0 : format!(
2301 0 : "failed to load remote timeline {} for tenant {}",
2302 0 : timeline_id, self.tenant_shard_id
2303 : )
2304 0 : })
2305 0 : .map_err(TimelineArchivalError::Other)?;
2306 :
2307 0 : let timeline = {
2308 0 : let timelines = self.timelines.lock().unwrap();
2309 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2310 0 : warn!("timeline not available directly after attach");
2311 : // This is not a panic because no locks are held between `load_remote_timeline`
2312 : // which puts the timeline into timelines, and our look into the timeline map.
2313 0 : return Err(TimelineArchivalError::Other(anyhow::anyhow!(
2314 0 : "timeline not available directly after attach"
2315 0 : )));
2316 : };
2317 0 : let mut offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2318 0 : match offloaded_timelines.remove(&timeline_id) {
2319 0 : Some(offloaded) => {
2320 0 : offloaded.delete_from_ancestor_with_timelines(&timelines);
2321 0 : }
2322 0 : None => warn!("timeline already removed from offloaded timelines"),
2323 : }
2324 :
2325 0 : self.initialize_gc_info(&timelines, &offloaded_timelines, Some(timeline_id));
2326 :
2327 0 : Arc::clone(timeline)
2328 : };
2329 :
2330 : // Upload new list of offloaded timelines to S3
2331 0 : self.maybe_upload_tenant_manifest().await?;
2332 :
2333 : // Activate the timeline (if it makes sense)
2334 0 : if !(timeline.is_broken() || timeline.is_stopping()) {
2335 0 : let background_jobs_can_start = None;
2336 0 : timeline.activate(
2337 0 : self.clone(),
2338 0 : broker_client.clone(),
2339 0 : background_jobs_can_start,
2340 0 : &ctx.with_scope_timeline(&timeline),
2341 0 : );
2342 0 : }
2343 :
2344 0 : info!("timeline unoffloading complete");
2345 0 : Ok(timeline)
2346 0 : }
2347 :
2348 0 : pub(crate) async fn apply_timeline_archival_config(
2349 0 : self: &Arc<Self>,
2350 0 : timeline_id: TimelineId,
2351 0 : new_state: TimelineArchivalState,
2352 0 : broker_client: storage_broker::BrokerClientChannel,
2353 0 : ctx: RequestContext,
2354 0 : ) -> Result<(), TimelineArchivalError> {
2355 0 : info!("setting timeline archival config");
2356 : // First part: figure out what is needed to do, and do validation
2357 0 : let timeline_or_unarchive_offloaded = 'outer: {
2358 0 : let timelines = self.timelines.lock().unwrap();
2359 :
2360 0 : let Some(timeline) = timelines.get(&timeline_id) else {
2361 0 : let offloaded_timelines = self.timelines_offloaded.lock().unwrap();
2362 0 : let Some(offloaded) = offloaded_timelines.get(&timeline_id) else {
2363 0 : return Err(TimelineArchivalError::NotFound);
2364 : };
2365 0 : if new_state == TimelineArchivalState::Archived {
2366 : // It's offloaded already, so nothing to do
2367 0 : return Ok(());
2368 0 : }
2369 0 : if let Some(ancestor_timeline_id) = offloaded.ancestor_timeline_id {
2370 0 : Self::check_ancestor_of_to_be_unarchived_is_not_archived(
2371 0 : ancestor_timeline_id,
2372 0 : &timelines,
2373 0 : &offloaded_timelines,
2374 0 : )?;
2375 0 : }
2376 0 : break 'outer None;
2377 : };
2378 :
2379 : // Do some validation. We release the timelines lock below, so there is potential
2380 : // for race conditions: these checks are more present to prevent misunderstandings of
2381 : // the API's capabilities, instead of serving as the sole way to defend their invariants.
2382 0 : match new_state {
2383 : TimelineArchivalState::Unarchived => {
2384 0 : Self::check_to_be_unarchived_timeline_has_no_archived_parent(timeline)?
2385 : }
2386 : TimelineArchivalState::Archived => {
2387 0 : Self::check_to_be_archived_has_no_unarchived_children(timeline_id, &timelines)?
2388 : }
2389 : }
2390 0 : Some(Arc::clone(timeline))
2391 : };
2392 :
2393 : // Second part: unoffload timeline (if needed)
2394 0 : let timeline = if let Some(timeline) = timeline_or_unarchive_offloaded {
2395 0 : timeline
2396 : } else {
2397 : // Turn offloaded timeline into a non-offloaded one
2398 0 : self.unoffload_timeline(timeline_id, broker_client, ctx)
2399 0 : .await?
2400 : };
2401 :
2402 : // Third part: upload new timeline archival state and block until it is present in S3
2403 0 : let upload_needed = match timeline
2404 0 : .remote_client
2405 0 : .schedule_index_upload_for_timeline_archival_state(new_state)
2406 : {
2407 0 : Ok(upload_needed) => upload_needed,
2408 0 : Err(e) => {
2409 0 : if timeline.cancel.is_cancelled() {
2410 0 : return Err(TimelineArchivalError::Cancelled);
2411 : } else {
2412 0 : return Err(TimelineArchivalError::Other(e));
2413 : }
2414 : }
2415 : };
2416 :
2417 0 : if upload_needed {
2418 0 : info!("Uploading new state");
2419 : const MAX_WAIT: Duration = Duration::from_secs(10);
2420 0 : let Ok(v) =
2421 0 : tokio::time::timeout(MAX_WAIT, timeline.remote_client.wait_completion()).await
2422 : else {
2423 0 : tracing::warn!("reached timeout for waiting on upload queue");
2424 0 : return Err(TimelineArchivalError::Timeout);
2425 : };
2426 0 : v.map_err(|e| match e {
2427 0 : WaitCompletionError::NotInitialized(e) => {
2428 0 : TimelineArchivalError::Other(anyhow::anyhow!(e))
2429 : }
2430 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2431 0 : TimelineArchivalError::Cancelled
2432 : }
2433 0 : })?;
2434 0 : }
2435 0 : Ok(())
2436 0 : }
2437 :
2438 1 : pub fn get_offloaded_timeline(
2439 1 : &self,
2440 1 : timeline_id: TimelineId,
2441 1 : ) -> Result<Arc<OffloadedTimeline>, GetTimelineError> {
2442 1 : self.timelines_offloaded
2443 1 : .lock()
2444 1 : .unwrap()
2445 1 : .get(&timeline_id)
2446 1 : .map(Arc::clone)
2447 1 : .ok_or(GetTimelineError::NotFound {
2448 1 : tenant_id: self.tenant_shard_id,
2449 1 : timeline_id,
2450 1 : })
2451 1 : }
2452 :
2453 2 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
2454 2 : self.tenant_shard_id
2455 2 : }
2456 :
2457 : /// Get Timeline handle for given Neon timeline ID.
2458 : /// This function is idempotent. It doesn't change internal state in any way.
2459 111 : pub fn get_timeline(
2460 111 : &self,
2461 111 : timeline_id: TimelineId,
2462 111 : active_only: bool,
2463 111 : ) -> Result<Arc<Timeline>, GetTimelineError> {
2464 111 : let timelines_accessor = self.timelines.lock().unwrap();
2465 111 : let timeline = timelines_accessor
2466 111 : .get(&timeline_id)
2467 111 : .ok_or(GetTimelineError::NotFound {
2468 111 : tenant_id: self.tenant_shard_id,
2469 111 : timeline_id,
2470 111 : })?;
2471 :
2472 110 : if active_only && !timeline.is_active() {
2473 0 : Err(GetTimelineError::NotActive {
2474 0 : tenant_id: self.tenant_shard_id,
2475 0 : timeline_id,
2476 0 : state: timeline.current_state(),
2477 0 : })
2478 : } else {
2479 110 : Ok(Arc::clone(timeline))
2480 : }
2481 111 : }
2482 :
2483 : /// Lists timelines the tenant contains.
2484 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2485 3 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
2486 3 : self.timelines
2487 3 : .lock()
2488 3 : .unwrap()
2489 3 : .values()
2490 3 : .map(Arc::clone)
2491 3 : .collect()
2492 3 : }
2493 :
2494 : /// Lists timelines the tenant contains.
2495 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2496 0 : pub fn list_importing_timelines(&self) -> Vec<Arc<ImportingTimeline>> {
2497 0 : self.timelines_importing
2498 0 : .lock()
2499 0 : .unwrap()
2500 0 : .values()
2501 0 : .map(Arc::clone)
2502 0 : .collect()
2503 0 : }
2504 :
2505 : /// Lists timelines the tenant manages, including offloaded ones.
2506 : ///
2507 : /// It's up to callers to omit certain timelines that are not considered ready for use.
2508 0 : pub fn list_timelines_and_offloaded(
2509 0 : &self,
2510 0 : ) -> (Vec<Arc<Timeline>>, Vec<Arc<OffloadedTimeline>>) {
2511 0 : let timelines = self
2512 0 : .timelines
2513 0 : .lock()
2514 0 : .unwrap()
2515 0 : .values()
2516 0 : .map(Arc::clone)
2517 0 : .collect();
2518 0 : let offloaded = self
2519 0 : .timelines_offloaded
2520 0 : .lock()
2521 0 : .unwrap()
2522 0 : .values()
2523 0 : .map(Arc::clone)
2524 0 : .collect();
2525 0 : (timelines, offloaded)
2526 0 : }
2527 :
2528 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
2529 0 : self.timelines.lock().unwrap().keys().cloned().collect()
2530 0 : }
2531 :
2532 : /// This is used by tests & import-from-basebackup.
2533 : ///
2534 : /// The returned [`UninitializedTimeline`] contains no data nor metadata and it is in
2535 : /// a state that will fail [`TenantShard::load_remote_timeline`] because `disk_consistent_lsn=Lsn(0)`.
2536 : ///
2537 : /// The caller is responsible for getting the timeline into a state that will be accepted
2538 : /// by [`TenantShard::load_remote_timeline`] / [`TenantShard::attach`].
2539 : /// Then they may call [`UninitializedTimeline::finish_creation`] to add the timeline
2540 : /// to the [`TenantShard::timelines`].
2541 : ///
2542 : /// Tests should use `TenantShard::create_test_timeline` to set up the minimum required metadata keys.
2543 114 : pub(crate) async fn create_empty_timeline(
2544 114 : self: &Arc<Self>,
2545 114 : new_timeline_id: TimelineId,
2546 114 : initdb_lsn: Lsn,
2547 114 : pg_version: PgMajorVersion,
2548 114 : ctx: &RequestContext,
2549 114 : ) -> anyhow::Result<(UninitializedTimeline, RequestContext)> {
2550 114 : anyhow::ensure!(
2551 114 : self.is_active(),
2552 0 : "Cannot create empty timelines on inactive tenant"
2553 : );
2554 :
2555 : // Protect against concurrent attempts to use this TimelineId
2556 114 : let create_guard = match self
2557 114 : .start_creating_timeline(new_timeline_id, CreateTimelineIdempotency::FailWithConflict)
2558 114 : .await?
2559 : {
2560 113 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2561 : StartCreatingTimelineResult::Idempotent(_) => {
2562 0 : unreachable!("FailWithConflict implies we get an error instead")
2563 : }
2564 : };
2565 :
2566 113 : let new_metadata = TimelineMetadata::new(
2567 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2568 : // make it valid, before calling finish_creation()
2569 113 : Lsn(0),
2570 113 : None,
2571 113 : None,
2572 113 : Lsn(0),
2573 113 : initdb_lsn,
2574 113 : initdb_lsn,
2575 113 : pg_version,
2576 : );
2577 113 : self.prepare_new_timeline(
2578 113 : new_timeline_id,
2579 113 : &new_metadata,
2580 113 : create_guard,
2581 113 : initdb_lsn,
2582 113 : None,
2583 113 : None,
2584 113 : ctx,
2585 113 : )
2586 113 : .await
2587 114 : }
2588 :
2589 : /// Helper for unit tests to create an empty timeline.
2590 : ///
2591 : /// The timeline is has state value `Active` but its background loops are not running.
2592 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
2593 : // Our current tests don't need the background loops.
2594 : #[cfg(test)]
2595 109 : pub async fn create_test_timeline(
2596 109 : self: &Arc<Self>,
2597 109 : new_timeline_id: TimelineId,
2598 109 : initdb_lsn: Lsn,
2599 109 : pg_version: PgMajorVersion,
2600 109 : ctx: &RequestContext,
2601 109 : ) -> anyhow::Result<Arc<Timeline>> {
2602 109 : let (uninit_tl, ctx) = self
2603 109 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2604 109 : .await?;
2605 109 : let tline = uninit_tl.raw_timeline().expect("we just created it");
2606 109 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
2607 :
2608 : // Setup minimum keys required for the timeline to be usable.
2609 109 : let mut modification = tline.begin_modification(initdb_lsn);
2610 109 : modification
2611 109 : .init_empty_test_timeline()
2612 109 : .context("init_empty_test_timeline")?;
2613 109 : modification
2614 109 : .commit(&ctx)
2615 109 : .await
2616 109 : .context("commit init_empty_test_timeline modification")?;
2617 :
2618 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
2619 109 : tline.maybe_spawn_flush_loop();
2620 109 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
2621 :
2622 : // Make sure the freeze_and_flush reaches remote storage.
2623 109 : tline.remote_client.wait_completion().await.unwrap();
2624 :
2625 109 : let tl = uninit_tl.finish_creation().await?;
2626 : // The non-test code would call tl.activate() here.
2627 109 : tl.set_state(TimelineState::Active);
2628 109 : Ok(tl)
2629 109 : }
2630 :
2631 : /// Helper for unit tests to create a timeline with some pre-loaded states.
2632 : #[cfg(test)]
2633 : #[allow(clippy::too_many_arguments)]
2634 24 : pub async fn create_test_timeline_with_layers(
2635 24 : self: &Arc<Self>,
2636 24 : new_timeline_id: TimelineId,
2637 24 : initdb_lsn: Lsn,
2638 24 : pg_version: PgMajorVersion,
2639 24 : ctx: &RequestContext,
2640 24 : in_memory_layer_desc: Vec<timeline::InMemoryLayerTestDesc>,
2641 24 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2642 24 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2643 24 : end_lsn: Lsn,
2644 24 : ) -> anyhow::Result<Arc<Timeline>> {
2645 : use checks::check_valid_layermap;
2646 : use itertools::Itertools;
2647 :
2648 24 : let tline = self
2649 24 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
2650 24 : .await?;
2651 24 : tline.force_advance_lsn(end_lsn);
2652 71 : for deltas in delta_layer_desc {
2653 47 : tline
2654 47 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
2655 47 : .await?;
2656 : }
2657 58 : for (lsn, images) in image_layer_desc {
2658 34 : tline
2659 34 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
2660 34 : .await?;
2661 : }
2662 28 : for in_memory in in_memory_layer_desc {
2663 4 : tline
2664 4 : .force_create_in_memory_layer(in_memory, Some(initdb_lsn), ctx)
2665 4 : .await?;
2666 : }
2667 24 : let layer_names = tline
2668 24 : .layers
2669 24 : .read(LayerManagerLockHolder::Testing)
2670 24 : .await
2671 24 : .layer_map()
2672 24 : .unwrap()
2673 24 : .iter_historic_layers()
2674 105 : .map(|layer| layer.layer_name())
2675 24 : .collect_vec();
2676 24 : if let Some(err) = check_valid_layermap(&layer_names) {
2677 0 : bail!("invalid layermap: {err}");
2678 24 : }
2679 24 : Ok(tline)
2680 24 : }
2681 :
2682 : /// Create a new timeline.
2683 : ///
2684 : /// Returns the new timeline ID and reference to its Timeline object.
2685 : ///
2686 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
2687 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
2688 : #[allow(clippy::too_many_arguments)]
2689 0 : pub(crate) async fn create_timeline(
2690 0 : self: &Arc<TenantShard>,
2691 0 : params: CreateTimelineParams,
2692 0 : broker_client: storage_broker::BrokerClientChannel,
2693 0 : ctx: &RequestContext,
2694 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2695 0 : if !self.is_active() {
2696 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
2697 0 : return Err(CreateTimelineError::ShuttingDown);
2698 : } else {
2699 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
2700 0 : "Cannot create timelines on inactive tenant"
2701 0 : )));
2702 : }
2703 0 : }
2704 :
2705 0 : let _gate = self
2706 0 : .gate
2707 0 : .enter()
2708 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
2709 :
2710 0 : let result: CreateTimelineResult = match params {
2711 : CreateTimelineParams::Bootstrap(CreateTimelineParamsBootstrap {
2712 0 : new_timeline_id,
2713 0 : existing_initdb_timeline_id,
2714 0 : pg_version,
2715 : }) => {
2716 0 : self.bootstrap_timeline(
2717 0 : new_timeline_id,
2718 0 : pg_version,
2719 0 : existing_initdb_timeline_id,
2720 0 : ctx,
2721 0 : )
2722 0 : .await?
2723 : }
2724 : CreateTimelineParams::Branch(CreateTimelineParamsBranch {
2725 0 : new_timeline_id,
2726 0 : ancestor_timeline_id,
2727 0 : mut ancestor_start_lsn,
2728 : }) => {
2729 0 : let ancestor_timeline = self
2730 0 : .get_timeline(ancestor_timeline_id, false)
2731 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
2732 :
2733 : // instead of waiting around, just deny the request because ancestor is not yet
2734 : // ready for other purposes either.
2735 0 : if !ancestor_timeline.is_active() {
2736 0 : return Err(CreateTimelineError::AncestorNotActive);
2737 0 : }
2738 :
2739 0 : if ancestor_timeline.is_archived() == Some(true) {
2740 0 : info!("tried to branch archived timeline");
2741 0 : return Err(CreateTimelineError::AncestorArchived);
2742 0 : }
2743 :
2744 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
2745 0 : *lsn = lsn.align();
2746 :
2747 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
2748 0 : if ancestor_ancestor_lsn > *lsn {
2749 : // can we safely just branch from the ancestor instead?
2750 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2751 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
2752 0 : lsn,
2753 0 : ancestor_timeline_id,
2754 0 : ancestor_ancestor_lsn,
2755 0 : )));
2756 0 : }
2757 :
2758 : // Wait for the WAL to arrive and be processed on the parent branch up
2759 : // to the requested branch point. The repository code itself doesn't
2760 : // require it, but if we start to receive WAL on the new timeline,
2761 : // decoding the new WAL might need to look up previous pages, relation
2762 : // sizes etc. and that would get confused if the previous page versions
2763 : // are not in the repository yet.
2764 0 : ancestor_timeline
2765 0 : .wait_lsn(
2766 0 : *lsn,
2767 0 : timeline::WaitLsnWaiter::Tenant,
2768 0 : timeline::WaitLsnTimeout::Default,
2769 0 : ctx,
2770 0 : )
2771 0 : .await
2772 0 : .map_err(|e| match e {
2773 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
2774 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
2775 : }
2776 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
2777 0 : })?;
2778 0 : }
2779 :
2780 0 : self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
2781 0 : .await?
2782 : }
2783 0 : CreateTimelineParams::ImportPgdata(params) => {
2784 0 : self.create_timeline_import_pgdata(params, ctx).await?
2785 : }
2786 : };
2787 :
2788 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
2789 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
2790 : // not send a success to the caller until it is. The same applies to idempotent retries.
2791 : //
2792 : // TODO: the timeline is already visible in [`Self::timelines`]; a caller could incorrectly
2793 : // assume that, because they can see the timeline via API, that the creation is done and
2794 : // that it is durable. Ideally, we would keep the timeline hidden (in [`Self::timelines_creating`])
2795 : // until it is durable, e.g., by extending the time we hold the creation guard. This also
2796 : // interacts with UninitializedTimeline and is generally a bit tricky.
2797 : //
2798 : // To re-emphasize: the only correct way to create a timeline is to repeat calling the
2799 : // creation API until it returns success. Only then is durability guaranteed.
2800 0 : info!(creation_result=%result.discriminant(), "waiting for timeline to be durable");
2801 0 : result
2802 0 : .timeline()
2803 0 : .remote_client
2804 0 : .wait_completion()
2805 0 : .await
2806 0 : .map_err(|e| match e {
2807 : WaitCompletionError::NotInitialized(
2808 0 : e, // If the queue is already stopped, it's a shutdown error.
2809 0 : ) if e.is_stopping() => CreateTimelineError::ShuttingDown,
2810 : WaitCompletionError::NotInitialized(_) => {
2811 : // This is a bug: we should never try to wait for uploads before initializing the timeline
2812 0 : debug_assert!(false);
2813 0 : CreateTimelineError::Other(anyhow::anyhow!("timeline not initialized"))
2814 : }
2815 : WaitCompletionError::UploadQueueShutDownOrStopped => {
2816 0 : CreateTimelineError::ShuttingDown
2817 : }
2818 0 : })?;
2819 :
2820 : // The creating task is responsible for activating the timeline.
2821 : // We do this after `wait_completion()` so that we don't spin up tasks that start
2822 : // doing stuff before the IndexPart is durable in S3, which is done by the previous section.
2823 0 : let activated_timeline = match result {
2824 0 : CreateTimelineResult::Created(timeline) => {
2825 0 : timeline.activate(
2826 0 : self.clone(),
2827 0 : broker_client,
2828 0 : None,
2829 0 : &ctx.with_scope_timeline(&timeline),
2830 : );
2831 0 : timeline
2832 : }
2833 0 : CreateTimelineResult::Idempotent(timeline) => {
2834 0 : info!(
2835 0 : "request was deemed idempotent, activation will be done by the creating task"
2836 : );
2837 0 : timeline
2838 : }
2839 0 : CreateTimelineResult::ImportSpawned(timeline) => {
2840 0 : info!(
2841 0 : "import task spawned, timeline will become visible and activated once the import is done"
2842 : );
2843 0 : timeline
2844 : }
2845 : };
2846 :
2847 0 : Ok(activated_timeline)
2848 0 : }
2849 :
2850 : /// The returned [`Arc<Timeline>`] is NOT in the [`TenantShard::timelines`] map until the import
2851 : /// completes in the background. A DIFFERENT [`Arc<Timeline>`] will be inserted into the
2852 : /// [`TenantShard::timelines`] map when the import completes.
2853 : /// We only return an [`Arc<Timeline>`] here so the API handler can create a [`pageserver_api::models::TimelineInfo`]
2854 : /// for the response.
2855 0 : async fn create_timeline_import_pgdata(
2856 0 : self: &Arc<Self>,
2857 0 : params: CreateTimelineParamsImportPgdata,
2858 0 : ctx: &RequestContext,
2859 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
2860 : let CreateTimelineParamsImportPgdata {
2861 0 : new_timeline_id,
2862 0 : location,
2863 0 : idempotency_key,
2864 0 : } = params;
2865 :
2866 0 : let started_at = chrono::Utc::now().naive_utc();
2867 :
2868 : //
2869 : // There's probably a simpler way to upload an index part, but, remote_timeline_client
2870 : // is the canonical way we do it.
2871 : // - create an empty timeline in-memory
2872 : // - use its remote_timeline_client to do the upload
2873 : // - dispose of the uninit timeline
2874 : // - keep the creation guard alive
2875 :
2876 0 : let timeline_create_guard = match self
2877 0 : .start_creating_timeline(
2878 0 : new_timeline_id,
2879 0 : CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
2880 0 : idempotency_key: idempotency_key.clone(),
2881 0 : }),
2882 0 : )
2883 0 : .await?
2884 : {
2885 0 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
2886 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
2887 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
2888 : }
2889 : };
2890 :
2891 0 : let (mut uninit_timeline, timeline_ctx) = {
2892 0 : let this = &self;
2893 0 : let initdb_lsn = Lsn(0);
2894 0 : async move {
2895 0 : let new_metadata = TimelineMetadata::new(
2896 : // Initialize disk_consistent LSN to 0, The caller must import some data to
2897 : // make it valid, before calling finish_creation()
2898 0 : Lsn(0),
2899 0 : None,
2900 0 : None,
2901 0 : Lsn(0),
2902 0 : initdb_lsn,
2903 0 : initdb_lsn,
2904 0 : PgMajorVersion::PG15,
2905 : );
2906 0 : this.prepare_new_timeline(
2907 0 : new_timeline_id,
2908 0 : &new_metadata,
2909 0 : timeline_create_guard,
2910 0 : initdb_lsn,
2911 0 : None,
2912 0 : None,
2913 0 : ctx,
2914 0 : )
2915 0 : .await
2916 0 : }
2917 : }
2918 0 : .await?;
2919 :
2920 0 : let in_progress = import_pgdata::index_part_format::InProgress {
2921 0 : idempotency_key,
2922 0 : location,
2923 0 : started_at,
2924 0 : };
2925 0 : let index_part = import_pgdata::index_part_format::Root::V1(
2926 0 : import_pgdata::index_part_format::V1::InProgress(in_progress),
2927 0 : );
2928 0 : uninit_timeline
2929 0 : .raw_timeline()
2930 0 : .unwrap()
2931 0 : .remote_client
2932 0 : .schedule_index_upload_for_import_pgdata_state_update(Some(index_part.clone()))?;
2933 :
2934 : // wait_completion happens in caller
2935 :
2936 0 : let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
2937 :
2938 0 : let import_task_gate = Gate::default();
2939 0 : let import_task_guard = import_task_gate.enter().unwrap();
2940 :
2941 0 : let import_task_handle = tokio::spawn(self.clone().create_timeline_import_pgdata_task(
2942 0 : timeline.clone(),
2943 0 : index_part,
2944 0 : timeline_create_guard,
2945 0 : import_task_guard,
2946 0 : timeline_ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
2947 : ));
2948 :
2949 0 : let prev = self.timelines_importing.lock().unwrap().insert(
2950 0 : timeline.timeline_id,
2951 0 : Arc::new(ImportingTimeline {
2952 0 : timeline: timeline.clone(),
2953 0 : import_task_handle,
2954 0 : import_task_gate,
2955 0 : delete_progress: TimelineDeleteProgress::default(),
2956 0 : }),
2957 0 : );
2958 :
2959 : // Idempotency is enforced higher up the stack
2960 0 : assert!(prev.is_none());
2961 :
2962 : // NB: the timeline doesn't exist in self.timelines at this point
2963 0 : Ok(CreateTimelineResult::ImportSpawned(timeline))
2964 0 : }
2965 :
2966 : /// Finalize the import of a timeline on this shard by marking it complete in
2967 : /// the index part. If the import task hasn't finished yet, returns an error.
2968 : ///
2969 : /// This method is idempotent. If the import was finalized once, the next call
2970 : /// will be a no-op.
2971 0 : pub(crate) async fn finalize_importing_timeline(
2972 0 : &self,
2973 0 : timeline_id: TimelineId,
2974 0 : ) -> Result<(), FinalizeTimelineImportError> {
2975 0 : let timeline = {
2976 0 : let locked = self.timelines_importing.lock().unwrap();
2977 0 : match locked.get(&timeline_id) {
2978 0 : Some(importing_timeline) => {
2979 0 : if !importing_timeline.import_task_handle.is_finished() {
2980 0 : return Err(FinalizeTimelineImportError::ImportTaskStillRunning);
2981 0 : }
2982 :
2983 0 : importing_timeline.timeline.clone()
2984 : }
2985 : None => {
2986 0 : return Ok(());
2987 : }
2988 : }
2989 : };
2990 :
2991 0 : timeline
2992 0 : .remote_client
2993 0 : .schedule_index_upload_for_import_pgdata_finalize()
2994 0 : .map_err(|_err| FinalizeTimelineImportError::ShuttingDown)?;
2995 0 : timeline
2996 0 : .remote_client
2997 0 : .wait_completion()
2998 0 : .await
2999 0 : .map_err(|_err| FinalizeTimelineImportError::ShuttingDown)?;
3000 :
3001 0 : self.timelines_importing
3002 0 : .lock()
3003 0 : .unwrap()
3004 0 : .remove(&timeline_id);
3005 :
3006 0 : Ok(())
3007 0 : }
3008 :
3009 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%timeline.timeline_id))]
3010 : async fn create_timeline_import_pgdata_task(
3011 : self: Arc<TenantShard>,
3012 : timeline: Arc<Timeline>,
3013 : index_part: import_pgdata::index_part_format::Root,
3014 : timeline_create_guard: TimelineCreateGuard,
3015 : _import_task_guard: GateGuard,
3016 : ctx: RequestContext,
3017 : ) {
3018 : debug_assert_current_span_has_tenant_and_timeline_id();
3019 : info!("starting");
3020 : scopeguard::defer! {info!("exiting")};
3021 :
3022 : let res = self
3023 : .create_timeline_import_pgdata_task_impl(
3024 : timeline,
3025 : index_part,
3026 : timeline_create_guard,
3027 : ctx,
3028 : )
3029 : .await;
3030 : if let Err(err) = &res {
3031 : error!(?err, "task failed");
3032 : // TODO sleep & retry, sensitive to tenant shutdown
3033 : // TODO: allow timeline deletion requests => should cancel the task
3034 : }
3035 : }
3036 :
3037 0 : async fn create_timeline_import_pgdata_task_impl(
3038 0 : self: Arc<TenantShard>,
3039 0 : timeline: Arc<Timeline>,
3040 0 : index_part: import_pgdata::index_part_format::Root,
3041 0 : _timeline_create_guard: TimelineCreateGuard,
3042 0 : ctx: RequestContext,
3043 0 : ) -> Result<(), anyhow::Error> {
3044 0 : info!("importing pgdata");
3045 0 : let ctx = ctx.with_scope_timeline(&timeline);
3046 0 : import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
3047 0 : .await
3048 0 : .context("import")?;
3049 0 : info!("import done - waiting for activation");
3050 :
3051 0 : anyhow::Ok(())
3052 0 : }
3053 :
3054 0 : pub(crate) async fn delete_timeline(
3055 0 : self: Arc<Self>,
3056 0 : timeline_id: TimelineId,
3057 0 : ) -> Result<(), DeleteTimelineError> {
3058 0 : DeleteTimelineFlow::run(&self, timeline_id).await?;
3059 :
3060 0 : Ok(())
3061 0 : }
3062 :
3063 : /// perform one garbage collection iteration, removing old data files from disk.
3064 : /// this function is periodically called by gc task.
3065 : /// also it can be explicitly requested through page server api 'do_gc' command.
3066 : ///
3067 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
3068 : ///
3069 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
3070 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
3071 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
3072 : /// `pitr` specifies the same as a time difference from the current time. The effective
3073 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
3074 : /// requires more history to be retained.
3075 : //
3076 377 : pub(crate) async fn gc_iteration(
3077 377 : &self,
3078 377 : target_timeline_id: Option<TimelineId>,
3079 377 : horizon: u64,
3080 377 : pitr: Duration,
3081 377 : cancel: &CancellationToken,
3082 377 : ctx: &RequestContext,
3083 377 : ) -> Result<GcResult, GcError> {
3084 : // Don't start doing work during shutdown
3085 377 : if let TenantState::Stopping { .. } = self.current_state() {
3086 0 : return Ok(GcResult::default());
3087 377 : }
3088 :
3089 : // there is a global allowed_error for this
3090 377 : if !self.is_active() {
3091 0 : return Err(GcError::NotActive);
3092 377 : }
3093 :
3094 : {
3095 377 : let conf = self.tenant_conf.load();
3096 :
3097 : // If we may not delete layers, then simply skip GC. Even though a tenant
3098 : // in AttachedMulti state could do GC and just enqueue the blocked deletions,
3099 : // the only advantage to doing it is to perhaps shrink the LayerMap metadata
3100 : // a bit sooner than we would achieve by waiting for AttachedSingle status.
3101 377 : if !conf.location.may_delete_layers_hint() {
3102 0 : info!("Skipping GC in location state {:?}", conf.location);
3103 0 : return Ok(GcResult::default());
3104 377 : }
3105 :
3106 377 : if conf.is_gc_blocked_by_lsn_lease_deadline() {
3107 0 : info!("Skipping GC because lsn lease deadline is not reached");
3108 0 : return Ok(GcResult::default());
3109 377 : }
3110 : }
3111 :
3112 377 : let _guard = match self.gc_block.start().await {
3113 377 : Ok(guard) => guard,
3114 0 : Err(reasons) => {
3115 0 : info!("Skipping GC: {reasons}");
3116 0 : return Ok(GcResult::default());
3117 : }
3118 : };
3119 :
3120 377 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
3121 377 : .await
3122 377 : }
3123 :
3124 : /// Performs one compaction iteration. Called periodically from the compaction loop. Returns
3125 : /// whether another compaction is needed, if we still have pending work or if we yield for
3126 : /// immediate L0 compaction.
3127 : ///
3128 : /// Compaction can also be explicitly requested for a timeline via the HTTP API.
3129 0 : async fn compaction_iteration(
3130 0 : self: &Arc<Self>,
3131 0 : cancel: &CancellationToken,
3132 0 : ctx: &RequestContext,
3133 0 : ) -> Result<CompactionOutcome, CompactionError> {
3134 : // Don't compact inactive tenants.
3135 0 : if !self.is_active() {
3136 0 : return Ok(CompactionOutcome::Skipped);
3137 0 : }
3138 :
3139 : // Don't compact tenants that can't upload layers. We don't check `may_delete_layers_hint`,
3140 : // since we need to compact L0 even in AttachedMulti to bound read amplification.
3141 0 : let location = self.tenant_conf.load().location;
3142 0 : if !location.may_upload_layers_hint() {
3143 0 : info!("skipping compaction in location state {location:?}");
3144 0 : return Ok(CompactionOutcome::Skipped);
3145 0 : }
3146 :
3147 : // Don't compact if the circuit breaker is tripped.
3148 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
3149 0 : info!("skipping compaction due to previous failures");
3150 0 : return Ok(CompactionOutcome::Skipped);
3151 0 : }
3152 :
3153 : // Collect all timelines to compact, along with offload instructions and L0 counts.
3154 0 : let mut compact: Vec<Arc<Timeline>> = Vec::new();
3155 0 : let mut offload: HashSet<TimelineId> = HashSet::new();
3156 0 : let mut l0_counts: HashMap<TimelineId, usize> = HashMap::new();
3157 :
3158 : {
3159 0 : let offload_enabled = self.get_timeline_offloading_enabled();
3160 0 : let timelines = self.timelines.lock().unwrap();
3161 0 : for (&timeline_id, timeline) in timelines.iter() {
3162 : // Skip inactive timelines.
3163 0 : if !timeline.is_active() {
3164 0 : continue;
3165 0 : }
3166 :
3167 : // Schedule the timeline for compaction.
3168 0 : compact.push(timeline.clone());
3169 :
3170 : // Schedule the timeline for offloading if eligible.
3171 0 : let can_offload = offload_enabled
3172 0 : && timeline.can_offload().0
3173 0 : && !timelines
3174 0 : .iter()
3175 0 : .any(|(_, tli)| tli.get_ancestor_timeline_id() == Some(timeline_id));
3176 0 : if can_offload {
3177 0 : offload.insert(timeline_id);
3178 0 : }
3179 : }
3180 : } // release timelines lock
3181 :
3182 0 : for timeline in &compact {
3183 : // Collect L0 counts. Can't await while holding lock above.
3184 0 : if let Ok(lm) = timeline
3185 0 : .layers
3186 0 : .read(LayerManagerLockHolder::Compaction)
3187 0 : .await
3188 0 : .layer_map()
3189 0 : {
3190 0 : l0_counts.insert(timeline.timeline_id, lm.level0_deltas().len());
3191 0 : }
3192 : }
3193 :
3194 : // Pass 1: L0 compaction across all timelines, in order of L0 count. We prioritize this to
3195 : // bound read amplification.
3196 : //
3197 : // TODO: this may spin on one or more ingest-heavy timelines, starving out image/GC
3198 : // compaction and offloading. We leave that as a potential problem to solve later. Consider
3199 : // splitting L0 and image/GC compaction to separate background jobs.
3200 0 : if self.get_compaction_l0_first() {
3201 0 : let compaction_threshold = self.get_compaction_threshold();
3202 0 : let compact_l0 = compact
3203 0 : .iter()
3204 0 : .map(|tli| (tli, l0_counts.get(&tli.timeline_id).copied().unwrap_or(0)))
3205 0 : .filter(|&(_, l0)| l0 >= compaction_threshold)
3206 0 : .sorted_by_key(|&(_, l0)| l0)
3207 0 : .rev()
3208 0 : .map(|(tli, _)| tli.clone())
3209 0 : .collect_vec();
3210 :
3211 0 : let mut has_pending_l0 = false;
3212 0 : for timeline in compact_l0 {
3213 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3214 : // NB: don't set CompactFlags::YieldForL0, since this is an L0-only compaction pass.
3215 0 : let outcome = timeline
3216 0 : .compact(cancel, CompactFlags::OnlyL0Compaction.into(), ctx)
3217 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3218 0 : .await
3219 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3220 0 : match outcome {
3221 0 : CompactionOutcome::Done => {}
3222 0 : CompactionOutcome::Skipped => {}
3223 0 : CompactionOutcome::Pending => has_pending_l0 = true,
3224 0 : CompactionOutcome::YieldForL0 => has_pending_l0 = true,
3225 : }
3226 : }
3227 0 : if has_pending_l0 {
3228 0 : return Ok(CompactionOutcome::YieldForL0); // do another pass
3229 0 : }
3230 0 : }
3231 :
3232 : // Pass 2: image compaction and timeline offloading. If any timelines have accumulated more
3233 : // L0 layers, they may also be compacted here. Image compaction will yield if there is
3234 : // pending L0 compaction on any tenant timeline.
3235 : //
3236 : // TODO: consider ordering timelines by some priority, e.g. time since last full compaction,
3237 : // amount of L1 delta debt or garbage, offload-eligible timelines first, etc.
3238 0 : let mut has_pending = false;
3239 0 : for timeline in compact {
3240 0 : if !timeline.is_active() {
3241 0 : continue;
3242 0 : }
3243 0 : let ctx = &ctx.with_scope_timeline(&timeline);
3244 :
3245 : // Yield for L0 if the separate L0 pass is enabled (otherwise there's no point).
3246 0 : let mut flags = EnumSet::default();
3247 0 : if self.get_compaction_l0_first() {
3248 0 : flags |= CompactFlags::YieldForL0;
3249 0 : }
3250 :
3251 0 : let mut outcome = timeline
3252 0 : .compact(cancel, flags, ctx)
3253 0 : .instrument(info_span!("compact_timeline", timeline_id = %timeline.timeline_id))
3254 0 : .await
3255 0 : .inspect_err(|err| self.maybe_trip_compaction_breaker(err))?;
3256 :
3257 : // If we're done compacting, check the scheduled GC compaction queue for more work.
3258 0 : if outcome == CompactionOutcome::Done {
3259 0 : let queue = {
3260 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3261 0 : guard
3262 0 : .entry(timeline.timeline_id)
3263 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()))
3264 0 : .clone()
3265 : };
3266 0 : let gc_compaction_strategy = self
3267 0 : .feature_resolver
3268 0 : .evaluate_multivariate("gc-comapction-strategy")
3269 0 : .ok();
3270 0 : let span = if let Some(gc_compaction_strategy) = gc_compaction_strategy {
3271 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id, strategy = %gc_compaction_strategy)
3272 : } else {
3273 0 : info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id)
3274 : };
3275 0 : outcome = queue
3276 0 : .iteration(cancel, ctx, &self.gc_block, &timeline)
3277 0 : .instrument(span)
3278 0 : .await?;
3279 0 : }
3280 :
3281 : // If we're done compacting, offload the timeline if requested.
3282 0 : if outcome == CompactionOutcome::Done && offload.contains(&timeline.timeline_id) {
3283 0 : pausable_failpoint!("before-timeline-auto-offload");
3284 0 : offload_timeline(self, &timeline)
3285 0 : .instrument(info_span!("offload_timeline", timeline_id = %timeline.timeline_id))
3286 0 : .await
3287 0 : .or_else(|err| match err {
3288 : // Ignore this, we likely raced with unarchival.
3289 0 : OffloadError::NotArchived => Ok(()),
3290 0 : OffloadError::AlreadyInProgress => Ok(()),
3291 0 : OffloadError::Cancelled => Err(CompactionError::ShuttingDown),
3292 : // don't break the anyhow chain
3293 0 : OffloadError::Other(err) => Err(CompactionError::Other(err)),
3294 0 : })?;
3295 0 : }
3296 :
3297 0 : match outcome {
3298 0 : CompactionOutcome::Done => {}
3299 0 : CompactionOutcome::Skipped => {}
3300 0 : CompactionOutcome::Pending => has_pending = true,
3301 : // This mostly makes sense when the L0-only pass above is enabled, since there's
3302 : // otherwise no guarantee that we'll start with the timeline that has high L0.
3303 0 : CompactionOutcome::YieldForL0 => return Ok(CompactionOutcome::YieldForL0),
3304 : }
3305 : }
3306 :
3307 : // Success! Untrip the breaker if necessary.
3308 0 : self.compaction_circuit_breaker
3309 0 : .lock()
3310 0 : .unwrap()
3311 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
3312 :
3313 0 : match has_pending {
3314 0 : true => Ok(CompactionOutcome::Pending),
3315 0 : false => Ok(CompactionOutcome::Done),
3316 : }
3317 0 : }
3318 :
3319 : /// Trips the compaction circuit breaker if appropriate.
3320 0 : pub(crate) fn maybe_trip_compaction_breaker(&self, err: &CompactionError) {
3321 0 : match err {
3322 0 : err if err.is_cancel() => {}
3323 0 : CompactionError::ShuttingDown => (),
3324 0 : CompactionError::CollectKeySpaceError(err) => {
3325 0 : // CollectKeySpaceError::Cancelled and PageRead::Cancelled are handled in `err.is_cancel` branch.
3326 0 : self.compaction_circuit_breaker
3327 0 : .lock()
3328 0 : .unwrap()
3329 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3330 0 : }
3331 0 : CompactionError::Other(err) => {
3332 0 : self.compaction_circuit_breaker
3333 0 : .lock()
3334 0 : .unwrap()
3335 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, err);
3336 0 : }
3337 : }
3338 0 : }
3339 :
3340 : /// Cancel scheduled compaction tasks
3341 0 : pub(crate) fn cancel_scheduled_compaction(&self, timeline_id: TimelineId) {
3342 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3343 0 : if let Some(q) = guard.get_mut(&timeline_id) {
3344 0 : q.cancel_scheduled();
3345 0 : }
3346 0 : }
3347 :
3348 0 : pub(crate) fn get_scheduled_compaction_tasks(
3349 0 : &self,
3350 0 : timeline_id: TimelineId,
3351 0 : ) -> Vec<CompactInfoResponse> {
3352 0 : let res = {
3353 0 : let guard = self.scheduled_compaction_tasks.lock().unwrap();
3354 0 : guard.get(&timeline_id).map(|q| q.remaining_jobs())
3355 : };
3356 0 : let Some((running, remaining)) = res else {
3357 0 : return Vec::new();
3358 : };
3359 0 : let mut result = Vec::new();
3360 0 : if let Some((id, running)) = running {
3361 0 : result.extend(running.into_compact_info_resp(id, true));
3362 0 : }
3363 0 : for (id, job) in remaining {
3364 0 : result.extend(job.into_compact_info_resp(id, false));
3365 0 : }
3366 0 : result
3367 0 : }
3368 :
3369 : /// Schedule a compaction task for a timeline.
3370 0 : pub(crate) async fn schedule_compaction(
3371 0 : &self,
3372 0 : timeline_id: TimelineId,
3373 0 : options: CompactOptions,
3374 0 : ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
3375 0 : let (tx, rx) = tokio::sync::oneshot::channel();
3376 0 : let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
3377 0 : let q = guard
3378 0 : .entry(timeline_id)
3379 0 : .or_insert_with(|| Arc::new(GcCompactionQueue::new()));
3380 0 : q.schedule_manual_compaction(options, Some(tx));
3381 0 : Ok(rx)
3382 0 : }
3383 :
3384 : /// Performs periodic housekeeping, via the tenant housekeeping background task.
3385 0 : async fn housekeeping(&self) {
3386 : // Call through to all timelines to freeze ephemeral layers as needed. This usually happens
3387 : // during ingest, but we don't want idle timelines to hold open layers for too long.
3388 : //
3389 : // We don't do this if the tenant can't upload layers (i.e. it's in stale attachment mode).
3390 : // We don't run compaction in this case either, and don't want to keep flushing tiny L0
3391 : // layers that won't be compacted down.
3392 0 : if self.tenant_conf.load().location.may_upload_layers_hint() {
3393 0 : let timelines = self
3394 0 : .timelines
3395 0 : .lock()
3396 0 : .unwrap()
3397 0 : .values()
3398 0 : .filter(|tli| tli.is_active())
3399 0 : .cloned()
3400 0 : .collect_vec();
3401 :
3402 0 : for timeline in timelines {
3403 0 : timeline.maybe_freeze_ephemeral_layer().await;
3404 : }
3405 0 : }
3406 :
3407 : // Shut down walredo if idle.
3408 : const WALREDO_IDLE_TIMEOUT: Duration = Duration::from_secs(180);
3409 0 : if let Some(ref walredo_mgr) = self.walredo_mgr {
3410 0 : walredo_mgr.maybe_quiesce(WALREDO_IDLE_TIMEOUT);
3411 0 : }
3412 :
3413 : // Update the feature resolver with the latest tenant-spcific data.
3414 0 : self.feature_resolver.refresh_properties_and_flags(self);
3415 0 : }
3416 :
3417 0 : pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
3418 0 : let timelines = self.timelines.lock().unwrap();
3419 0 : !timelines
3420 0 : .iter()
3421 0 : .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(timeline_id))
3422 0 : }
3423 :
3424 1369 : pub fn current_state(&self) -> TenantState {
3425 1369 : self.state.borrow().clone()
3426 1369 : }
3427 :
3428 988 : pub fn is_active(&self) -> bool {
3429 988 : self.current_state() == TenantState::Active
3430 988 : }
3431 :
3432 0 : pub fn generation(&self) -> Generation {
3433 0 : self.generation
3434 0 : }
3435 :
3436 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
3437 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
3438 0 : }
3439 :
3440 : /// Changes tenant status to active, unless shutdown was already requested.
3441 : ///
3442 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
3443 : /// to delay background jobs. Background jobs can be started right away when None is given.
3444 0 : fn activate(
3445 0 : self: &Arc<Self>,
3446 0 : broker_client: BrokerClientChannel,
3447 0 : background_jobs_can_start: Option<&completion::Barrier>,
3448 0 : ctx: &RequestContext,
3449 0 : ) {
3450 0 : span::debug_assert_current_span_has_tenant_id();
3451 :
3452 0 : let mut activating = false;
3453 0 : self.state.send_modify(|current_state| {
3454 : use pageserver_api::models::ActivatingFrom;
3455 0 : match &*current_state {
3456 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
3457 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {current_state:?}");
3458 : }
3459 0 : TenantState::Attaching => {
3460 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
3461 0 : }
3462 : }
3463 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
3464 0 : activating = true;
3465 : // Continue outside the closure. We need to grab timelines.lock()
3466 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3467 0 : });
3468 :
3469 0 : if activating {
3470 0 : let timelines_accessor = self.timelines.lock().unwrap();
3471 0 : let timelines_offloaded_accessor = self.timelines_offloaded.lock().unwrap();
3472 0 : let timelines_to_activate = timelines_accessor
3473 0 : .values()
3474 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
3475 :
3476 : // Spawn gc and compaction loops. The loops will shut themselves
3477 : // down when they notice that the tenant is inactive.
3478 0 : tasks::start_background_loops(self, background_jobs_can_start);
3479 :
3480 0 : let mut activated_timelines = 0;
3481 :
3482 0 : for timeline in timelines_to_activate {
3483 0 : timeline.activate(
3484 0 : self.clone(),
3485 0 : broker_client.clone(),
3486 0 : background_jobs_can_start,
3487 0 : &ctx.with_scope_timeline(timeline),
3488 0 : );
3489 0 : activated_timelines += 1;
3490 0 : }
3491 :
3492 0 : let tid = self.tenant_shard_id.tenant_id.to_string();
3493 0 : let shard_id = self.tenant_shard_id.shard_slug().to_string();
3494 0 : let offloaded_timeline_count = timelines_offloaded_accessor.len();
3495 0 : TENANT_OFFLOADED_TIMELINES
3496 0 : .with_label_values(&[&tid, &shard_id])
3497 0 : .set(offloaded_timeline_count as u64);
3498 :
3499 0 : self.state.send_modify(move |current_state| {
3500 0 : assert!(
3501 0 : matches!(current_state, TenantState::Activating(_)),
3502 0 : "set_stopping and set_broken wait for us to leave Activating state",
3503 : );
3504 0 : *current_state = TenantState::Active;
3505 :
3506 0 : let elapsed = self.constructed_at.elapsed();
3507 0 : let total_timelines = timelines_accessor.len();
3508 :
3509 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
3510 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
3511 0 : info!(
3512 0 : since_creation_millis = elapsed.as_millis(),
3513 0 : tenant_id = %self.tenant_shard_id.tenant_id,
3514 0 : shard_id = %self.tenant_shard_id.shard_slug(),
3515 : activated_timelines,
3516 : total_timelines,
3517 0 : post_state = <&'static str>::from(&*current_state),
3518 0 : "activation attempt finished"
3519 : );
3520 :
3521 0 : TENANT.activation.observe(elapsed.as_secs_f64());
3522 0 : });
3523 0 : }
3524 0 : }
3525 :
3526 : /// Shutdown the tenant and join all of the spawned tasks.
3527 : ///
3528 : /// The method caters for all use-cases:
3529 : /// - pageserver shutdown (freeze_and_flush == true)
3530 : /// - detach + ignore (freeze_and_flush == false)
3531 : ///
3532 : /// This will attempt to shutdown even if tenant is broken.
3533 : ///
3534 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
3535 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
3536 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
3537 : /// the ongoing shutdown.
3538 3 : async fn shutdown(
3539 3 : &self,
3540 3 : shutdown_progress: completion::Barrier,
3541 3 : shutdown_mode: timeline::ShutdownMode,
3542 3 : ) -> Result<(), completion::Barrier> {
3543 3 : span::debug_assert_current_span_has_tenant_id();
3544 :
3545 : // Set tenant (and its timlines) to Stoppping state.
3546 : //
3547 : // Since we can only transition into Stopping state after activation is complete,
3548 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
3549 : //
3550 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
3551 : // 1. Lock out any new requests to the tenants.
3552 : // 2. Signal cancellation to WAL receivers (we wait on it below).
3553 : // 3. Signal cancellation for other tenant background loops.
3554 : // 4. ???
3555 : //
3556 : // The waiting for the cancellation is not done uniformly.
3557 : // We certainly wait for WAL receivers to shut down.
3558 : // That is necessary so that no new data comes in before the freeze_and_flush.
3559 : // But the tenant background loops are joined-on in our caller.
3560 : // It's mesed up.
3561 : // we just ignore the failure to stop
3562 :
3563 : // If we're still attaching, fire the cancellation token early to drop out: this
3564 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
3565 : // is very slow.
3566 3 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
3567 0 : self.cancel.cancel();
3568 :
3569 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
3570 : // are children of ours, so their flush loops will have shut down already
3571 0 : timeline::ShutdownMode::Hard
3572 : } else {
3573 3 : shutdown_mode
3574 : };
3575 :
3576 3 : match self.set_stopping(shutdown_progress).await {
3577 3 : Ok(()) => {}
3578 0 : Err(SetStoppingError::Broken) => {
3579 0 : // assume that this is acceptable
3580 0 : }
3581 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
3582 : // give caller the option to wait for this this shutdown
3583 0 : info!("Tenant::shutdown: AlreadyStopping");
3584 0 : return Err(other);
3585 : }
3586 : };
3587 :
3588 3 : let mut js = tokio::task::JoinSet::new();
3589 : {
3590 3 : let timelines = self.timelines.lock().unwrap();
3591 3 : timelines.values().for_each(|timeline| {
3592 3 : let timeline = Arc::clone(timeline);
3593 3 : let timeline_id = timeline.timeline_id;
3594 3 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
3595 3 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
3596 3 : });
3597 : }
3598 : {
3599 3 : let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
3600 3 : timelines_offloaded.values().for_each(|timeline| {
3601 0 : timeline.defuse_for_tenant_drop();
3602 0 : });
3603 : }
3604 : {
3605 3 : let mut timelines_importing = self.timelines_importing.lock().unwrap();
3606 3 : timelines_importing
3607 3 : .drain()
3608 3 : .for_each(|(timeline_id, importing_timeline)| {
3609 0 : let span = tracing::info_span!("importing_timeline_shutdown", %timeline_id);
3610 0 : js.spawn(async move { importing_timeline.shutdown().instrument(span).await });
3611 0 : });
3612 : }
3613 : // test_long_timeline_create_then_tenant_delete is leaning on this message
3614 3 : tracing::info!("Waiting for timelines...");
3615 6 : while let Some(res) = js.join_next().await {
3616 0 : match res {
3617 3 : Ok(()) => {}
3618 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
3619 0 : Err(je) if je.is_panic() => { /* logged already */ }
3620 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
3621 : }
3622 : }
3623 :
3624 3 : if let ShutdownMode::Reload = shutdown_mode {
3625 0 : tracing::info!("Flushing deletion queue");
3626 0 : if let Err(e) = self.deletion_queue_client.flush().await {
3627 0 : match e {
3628 0 : DeletionQueueError::ShuttingDown => {
3629 0 : // This is the only error we expect for now. In the future, if more error
3630 0 : // variants are added, we should handle them here.
3631 0 : }
3632 : }
3633 0 : }
3634 3 : }
3635 :
3636 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
3637 : // them to continue to do work during their shutdown methods, e.g. flushing data.
3638 3 : tracing::debug!("Cancelling CancellationToken");
3639 3 : self.cancel.cancel();
3640 :
3641 : // shutdown all tenant and timeline tasks: gc, compaction, page service
3642 : // No new tasks will be started for this tenant because it's in `Stopping` state.
3643 : //
3644 : // this will additionally shutdown and await all timeline tasks.
3645 3 : tracing::debug!("Waiting for tasks...");
3646 3 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
3647 :
3648 3 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
3649 3 : walredo_mgr.shutdown().await;
3650 0 : }
3651 :
3652 : // Wait for any in-flight operations to complete
3653 3 : self.gate.close().await;
3654 :
3655 3 : remove_tenant_metrics(&self.tenant_shard_id);
3656 :
3657 3 : Ok(())
3658 3 : }
3659 :
3660 : /// Change tenant status to Stopping, to mark that it is being shut down.
3661 : ///
3662 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3663 : ///
3664 : /// This function is not cancel-safe!
3665 3 : async fn set_stopping(&self, progress: completion::Barrier) -> Result<(), SetStoppingError> {
3666 3 : let mut rx = self.state.subscribe();
3667 :
3668 : // cannot stop before we're done activating, so wait out until we're done activating
3669 3 : rx.wait_for(|state| match state {
3670 : TenantState::Activating(_) | TenantState::Attaching => {
3671 0 : info!("waiting for {state} to turn Active|Broken|Stopping");
3672 0 : false
3673 : }
3674 3 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3675 3 : })
3676 3 : .await
3677 3 : .expect("cannot drop self.state while on a &self method");
3678 :
3679 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
3680 3 : let mut err = None;
3681 3 : let stopping = self.state.send_if_modified(|current_state| match current_state {
3682 : TenantState::Activating(_) | TenantState::Attaching => {
3683 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3684 : }
3685 : TenantState::Active => {
3686 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
3687 : // are created after the transition to Stopping. That's harmless, as the Timelines
3688 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
3689 3 : *current_state = TenantState::Stopping { progress: Some(progress) };
3690 : // Continue stopping outside the closure. We need to grab timelines.lock()
3691 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
3692 3 : true
3693 : }
3694 : TenantState::Stopping { progress: None } => {
3695 : // An attach was cancelled, and the attach transitioned the tenant from Attaching to
3696 : // Stopping(None) to let us know it exited. Register our progress and continue.
3697 0 : *current_state = TenantState::Stopping { progress: Some(progress) };
3698 0 : true
3699 : }
3700 0 : TenantState::Broken { reason, .. } => {
3701 0 : info!(
3702 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
3703 : );
3704 0 : err = Some(SetStoppingError::Broken);
3705 0 : false
3706 : }
3707 0 : TenantState::Stopping { progress: Some(progress) } => {
3708 0 : info!("Tenant is already in Stopping state");
3709 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
3710 0 : false
3711 : }
3712 3 : });
3713 3 : match (stopping, err) {
3714 3 : (true, None) => {} // continue
3715 0 : (false, Some(err)) => return Err(err),
3716 0 : (true, Some(_)) => unreachable!(
3717 : "send_if_modified closure must error out if not transitioning to Stopping"
3718 : ),
3719 0 : (false, None) => unreachable!(
3720 : "send_if_modified closure must return true if transitioning to Stopping"
3721 : ),
3722 : }
3723 :
3724 3 : let timelines_accessor = self.timelines.lock().unwrap();
3725 3 : let not_broken_timelines = timelines_accessor
3726 3 : .values()
3727 3 : .filter(|timeline| !timeline.is_broken());
3728 6 : for timeline in not_broken_timelines {
3729 3 : timeline.set_state(TimelineState::Stopping);
3730 3 : }
3731 3 : Ok(())
3732 3 : }
3733 :
3734 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
3735 : /// `remove_tenant_from_memory`
3736 : ///
3737 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
3738 : ///
3739 : /// In tests, we also use this to set tenants to Broken state on purpose.
3740 0 : pub(crate) async fn set_broken(&self, reason: String) {
3741 0 : let mut rx = self.state.subscribe();
3742 :
3743 : // The load & attach routines own the tenant state until it has reached `Active`.
3744 : // So, wait until it's done.
3745 0 : rx.wait_for(|state| match state {
3746 : TenantState::Activating(_) | TenantState::Attaching => {
3747 0 : info!(
3748 0 : "waiting for {} to turn Active|Broken|Stopping",
3749 0 : <&'static str>::from(state)
3750 : );
3751 0 : false
3752 : }
3753 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
3754 0 : })
3755 0 : .await
3756 0 : .expect("cannot drop self.state while on a &self method");
3757 :
3758 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
3759 0 : self.set_broken_no_wait(reason)
3760 0 : }
3761 :
3762 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
3763 0 : let reason = reason.to_string();
3764 0 : self.state.send_modify(|current_state| {
3765 0 : match *current_state {
3766 : TenantState::Activating(_) | TenantState::Attaching => {
3767 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
3768 : }
3769 : TenantState::Active => {
3770 0 : if cfg!(feature = "testing") {
3771 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
3772 0 : *current_state = TenantState::broken_from_reason(reason);
3773 : } else {
3774 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
3775 : }
3776 : }
3777 : TenantState::Broken { .. } => {
3778 0 : warn!("Tenant is already in Broken state");
3779 : }
3780 : // This is the only "expected" path, any other path is a bug.
3781 : TenantState::Stopping { .. } => {
3782 0 : warn!(
3783 0 : "Marking Stopping tenant as Broken state, reason: {}",
3784 : reason
3785 : );
3786 0 : *current_state = TenantState::broken_from_reason(reason);
3787 : }
3788 : }
3789 0 : });
3790 0 : }
3791 :
3792 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
3793 0 : self.state.subscribe()
3794 0 : }
3795 :
3796 : /// The activate_now semaphore is initialized with zero units. As soon as
3797 : /// we add a unit, waiters will be able to acquire a unit and proceed.
3798 0 : pub(crate) fn activate_now(&self) {
3799 0 : self.activate_now_sem.add_permits(1);
3800 0 : }
3801 :
3802 0 : pub(crate) async fn wait_to_become_active(
3803 0 : &self,
3804 0 : timeout: Duration,
3805 0 : ) -> Result<(), GetActiveTenantError> {
3806 0 : let mut receiver = self.state.subscribe();
3807 : loop {
3808 0 : let current_state = receiver.borrow_and_update().clone();
3809 0 : match current_state {
3810 : TenantState::Attaching | TenantState::Activating(_) => {
3811 : // in these states, there's a chance that we can reach ::Active
3812 0 : self.activate_now();
3813 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
3814 0 : Ok(r) => {
3815 0 : r.map_err(
3816 : |_e: tokio::sync::watch::error::RecvError|
3817 : // Tenant existed but was dropped: report it as non-existent
3818 0 : GetActiveTenantError::NotFound(GetTenantError::ShardNotFound(self.tenant_shard_id))
3819 0 : )?
3820 : }
3821 : Err(TimeoutCancellableError::Cancelled) => {
3822 0 : return Err(GetActiveTenantError::Cancelled);
3823 : }
3824 : Err(TimeoutCancellableError::Timeout) => {
3825 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
3826 0 : latest_state: Some(self.current_state()),
3827 0 : wait_time: timeout,
3828 0 : });
3829 : }
3830 : }
3831 : }
3832 : TenantState::Active => {
3833 0 : return Ok(());
3834 : }
3835 0 : TenantState::Broken { reason, .. } => {
3836 : // This is fatal, and reported distinctly from the general case of "will never be active" because
3837 : // it's logically a 500 to external API users (broken is always a bug).
3838 0 : return Err(GetActiveTenantError::Broken(reason));
3839 : }
3840 : TenantState::Stopping { .. } => {
3841 : // There's no chance the tenant can transition back into ::Active
3842 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
3843 : }
3844 : }
3845 : }
3846 0 : }
3847 :
3848 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
3849 0 : self.tenant_conf.load().location.attach_mode
3850 0 : }
3851 :
3852 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
3853 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
3854 : /// rare external API calls, like a reconciliation at startup.
3855 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
3856 0 : let attached_tenant_conf = self.tenant_conf.load();
3857 :
3858 0 : let location_config_mode = match attached_tenant_conf.location.attach_mode {
3859 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
3860 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
3861 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
3862 : };
3863 :
3864 0 : models::LocationConfig {
3865 0 : mode: location_config_mode,
3866 0 : generation: self.generation.into(),
3867 0 : secondary_conf: None,
3868 0 : shard_number: self.shard_identity.number.0,
3869 0 : shard_count: self.shard_identity.count.literal(),
3870 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
3871 0 : tenant_conf: attached_tenant_conf.tenant_conf.clone(),
3872 0 : }
3873 0 : }
3874 :
3875 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
3876 0 : &self.tenant_shard_id
3877 0 : }
3878 :
3879 0 : pub(crate) fn get_shard_identity(&self) -> ShardIdentity {
3880 0 : self.shard_identity
3881 0 : }
3882 :
3883 119 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
3884 119 : self.shard_identity.stripe_size
3885 119 : }
3886 :
3887 0 : pub(crate) fn get_generation(&self) -> Generation {
3888 0 : self.generation
3889 0 : }
3890 :
3891 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
3892 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
3893 : /// resetting this tenant to a valid state if we fail.
3894 0 : pub(crate) async fn split_prepare(
3895 0 : &self,
3896 0 : child_shards: &Vec<TenantShardId>,
3897 0 : ) -> anyhow::Result<()> {
3898 0 : let (timelines, offloaded) = {
3899 0 : let timelines = self.timelines.lock().unwrap();
3900 0 : let offloaded = self.timelines_offloaded.lock().unwrap();
3901 0 : (timelines.clone(), offloaded.clone())
3902 0 : };
3903 0 : let timelines_iter = timelines
3904 0 : .values()
3905 0 : .map(TimelineOrOffloadedArcRef::<'_>::from)
3906 0 : .chain(
3907 0 : offloaded
3908 0 : .values()
3909 0 : .map(TimelineOrOffloadedArcRef::<'_>::from),
3910 : );
3911 0 : for timeline in timelines_iter {
3912 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
3913 : // to ensure that they do not start a split if currently in the process of doing these.
3914 :
3915 0 : let timeline_id = timeline.timeline_id();
3916 :
3917 0 : if let TimelineOrOffloadedArcRef::Timeline(timeline) = timeline {
3918 : // Upload an index from the parent: this is partly to provide freshness for the
3919 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
3920 : // always be a parent shard index in the same generation as we wrote the child shard index.
3921 0 : tracing::info!(%timeline_id, "Uploading index");
3922 0 : timeline
3923 0 : .remote_client
3924 0 : .schedule_index_upload_for_file_changes()?;
3925 0 : timeline.remote_client.wait_completion().await?;
3926 0 : }
3927 :
3928 0 : let remote_client = match timeline {
3929 0 : TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.remote_client.clone(),
3930 0 : TimelineOrOffloadedArcRef::Offloaded(offloaded) => {
3931 0 : let remote_client = self
3932 0 : .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
3933 0 : Arc::new(remote_client)
3934 : }
3935 : TimelineOrOffloadedArcRef::Importing(_) => {
3936 0 : unreachable!("Importing timelines are not included in the iterator")
3937 : }
3938 : };
3939 :
3940 : // Shut down the timeline's remote client: this means that the indices we write
3941 : // for child shards will not be invalidated by the parent shard deleting layers.
3942 0 : tracing::info!(%timeline_id, "Shutting down remote storage client");
3943 0 : remote_client.shutdown().await;
3944 :
3945 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
3946 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
3947 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
3948 : // we use here really is the remotely persistent one).
3949 0 : tracing::info!(%timeline_id, "Downloading index_part from parent");
3950 0 : let result = remote_client
3951 0 : .download_index_file(&self.cancel)
3952 0 : .instrument(info_span!("download_index_file", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))
3953 0 : .await?;
3954 0 : let index_part = match result {
3955 : MaybeDeletedIndexPart::Deleted(_) => {
3956 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
3957 : }
3958 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
3959 : };
3960 :
3961 : // A shard split may not take place while a timeline import is on-going
3962 : // for the tenant. Timeline imports run as part of each tenant shard
3963 : // and rely on the sharding scheme to split the work among pageservers.
3964 : // If we were to split in the middle of this process, we would have to
3965 : // either ensure that it's driven to completion on the old shard set
3966 : // or transfer it to the new shard set. It's technically possible, but complex.
3967 0 : match index_part.import_pgdata {
3968 0 : Some(ref import) if !import.is_done() => {
3969 0 : anyhow::bail!(
3970 0 : "Cannot split due to import with idempotency key: {:?}",
3971 0 : import.idempotency_key()
3972 : );
3973 : }
3974 0 : Some(_) | None => {
3975 0 : // fallthrough
3976 0 : }
3977 : }
3978 :
3979 0 : for child_shard in child_shards {
3980 0 : tracing::info!(%timeline_id, "Uploading index_part for child {}", child_shard.to_index());
3981 0 : upload_index_part(
3982 0 : &self.remote_storage,
3983 0 : child_shard,
3984 0 : &timeline_id,
3985 0 : self.generation,
3986 0 : &index_part,
3987 0 : &self.cancel,
3988 0 : )
3989 0 : .await?;
3990 : }
3991 : }
3992 :
3993 0 : let tenant_manifest = self.build_tenant_manifest();
3994 0 : for child_shard in child_shards {
3995 0 : tracing::info!(
3996 0 : "Uploading tenant manifest for child {}",
3997 0 : child_shard.to_index()
3998 : );
3999 0 : upload_tenant_manifest(
4000 0 : &self.remote_storage,
4001 0 : child_shard,
4002 0 : self.generation,
4003 0 : &tenant_manifest,
4004 0 : &self.cancel,
4005 0 : )
4006 0 : .await?;
4007 : }
4008 :
4009 0 : Ok(())
4010 0 : }
4011 :
4012 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
4013 0 : let mut result = TopTenantShardItem {
4014 0 : id: self.tenant_shard_id,
4015 0 : resident_size: 0,
4016 0 : physical_size: 0,
4017 0 : max_logical_size: 0,
4018 0 : max_logical_size_per_shard: 0,
4019 0 : };
4020 :
4021 0 : for timeline in self.timelines.lock().unwrap().values() {
4022 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
4023 0 :
4024 0 : result.physical_size += timeline
4025 0 : .remote_client
4026 0 : .metrics
4027 0 : .remote_physical_size_gauge
4028 0 : .get();
4029 0 : result.max_logical_size = std::cmp::max(
4030 0 : result.max_logical_size,
4031 0 : timeline.metrics.current_logical_size_gauge.get(),
4032 0 : );
4033 0 : }
4034 :
4035 0 : result.max_logical_size_per_shard = result
4036 0 : .max_logical_size
4037 0 : .div_ceil(self.tenant_shard_id.shard_count.count() as u64);
4038 :
4039 0 : result
4040 0 : }
4041 : }
4042 :
4043 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
4044 : /// perform a topological sort, so that the parent of each timeline comes
4045 : /// before the children.
4046 : /// E extracts the ancestor from T
4047 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
4048 118 : fn tree_sort_timelines<T, E>(
4049 118 : timelines: HashMap<TimelineId, T>,
4050 118 : extractor: E,
4051 118 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
4052 118 : where
4053 118 : E: Fn(&T) -> Option<TimelineId>,
4054 : {
4055 118 : let mut result = Vec::with_capacity(timelines.len());
4056 :
4057 118 : let mut now = Vec::with_capacity(timelines.len());
4058 : // (ancestor, children)
4059 118 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
4060 118 : HashMap::with_capacity(timelines.len());
4061 :
4062 121 : for (timeline_id, value) in timelines {
4063 3 : if let Some(ancestor_id) = extractor(&value) {
4064 1 : let children = later.entry(ancestor_id).or_default();
4065 1 : children.push((timeline_id, value));
4066 2 : } else {
4067 2 : now.push((timeline_id, value));
4068 2 : }
4069 : }
4070 :
4071 121 : while let Some((timeline_id, metadata)) = now.pop() {
4072 3 : result.push((timeline_id, metadata));
4073 : // All children of this can be loaded now
4074 3 : if let Some(mut children) = later.remove(&timeline_id) {
4075 1 : now.append(&mut children);
4076 2 : }
4077 : }
4078 :
4079 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
4080 118 : if !later.is_empty() {
4081 0 : for (missing_id, orphan_ids) in later {
4082 0 : for (orphan_id, _) in orphan_ids {
4083 0 : error!(
4084 0 : "could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded"
4085 : );
4086 : }
4087 : }
4088 0 : bail!("could not load tenant because some timelines are missing ancestors");
4089 118 : }
4090 :
4091 118 : Ok(result)
4092 118 : }
4093 :
4094 : impl TenantShard {
4095 0 : pub fn tenant_specific_overrides(&self) -> pageserver_api::models::TenantConfig {
4096 0 : self.tenant_conf.load().tenant_conf.clone()
4097 0 : }
4098 :
4099 0 : pub fn effective_config(&self) -> pageserver_api::config::TenantConfigToml {
4100 0 : self.tenant_specific_overrides()
4101 0 : .merge(self.conf.default_tenant_conf.clone())
4102 0 : }
4103 :
4104 0 : pub fn get_checkpoint_distance(&self) -> u64 {
4105 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4106 0 : tenant_conf
4107 0 : .checkpoint_distance
4108 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
4109 0 : }
4110 :
4111 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
4112 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4113 0 : tenant_conf
4114 0 : .checkpoint_timeout
4115 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
4116 0 : }
4117 :
4118 0 : pub fn get_compaction_target_size(&self) -> u64 {
4119 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4120 0 : tenant_conf
4121 0 : .compaction_target_size
4122 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
4123 0 : }
4124 :
4125 0 : pub fn get_compaction_period(&self) -> Duration {
4126 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4127 0 : tenant_conf
4128 0 : .compaction_period
4129 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
4130 0 : }
4131 :
4132 0 : pub fn get_compaction_threshold(&self) -> usize {
4133 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4134 0 : tenant_conf
4135 0 : .compaction_threshold
4136 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
4137 0 : }
4138 :
4139 0 : pub fn get_rel_size_v2_enabled(&self) -> bool {
4140 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4141 0 : tenant_conf
4142 0 : .rel_size_v2_enabled
4143 0 : .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
4144 0 : }
4145 :
4146 0 : pub fn get_compaction_upper_limit(&self) -> usize {
4147 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4148 0 : tenant_conf
4149 0 : .compaction_upper_limit
4150 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
4151 0 : }
4152 :
4153 0 : pub fn get_compaction_l0_first(&self) -> bool {
4154 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4155 0 : tenant_conf
4156 0 : .compaction_l0_first
4157 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_l0_first)
4158 0 : }
4159 :
4160 120 : pub fn get_gc_horizon(&self) -> u64 {
4161 120 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4162 120 : tenant_conf
4163 120 : .gc_horizon
4164 120 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
4165 120 : }
4166 :
4167 0 : pub fn get_gc_period(&self) -> Duration {
4168 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4169 0 : tenant_conf
4170 0 : .gc_period
4171 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
4172 0 : }
4173 :
4174 0 : pub fn get_image_creation_threshold(&self) -> usize {
4175 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4176 0 : tenant_conf
4177 0 : .image_creation_threshold
4178 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
4179 0 : }
4180 :
4181 2 : pub fn get_pitr_interval(&self) -> Duration {
4182 2 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4183 2 : tenant_conf
4184 2 : .pitr_interval
4185 2 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
4186 2 : }
4187 :
4188 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
4189 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4190 0 : tenant_conf
4191 0 : .min_resident_size_override
4192 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
4193 0 : }
4194 :
4195 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
4196 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4197 0 : let heatmap_period = tenant_conf
4198 0 : .heatmap_period
4199 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
4200 0 : if heatmap_period.is_zero() {
4201 0 : None
4202 : } else {
4203 0 : Some(heatmap_period)
4204 : }
4205 0 : }
4206 :
4207 0 : pub fn get_lsn_lease_length(&self) -> Duration {
4208 0 : Self::get_lsn_lease_length_impl(self.conf, &self.tenant_conf.load().tenant_conf)
4209 0 : }
4210 :
4211 118 : pub fn get_lsn_lease_length_impl(
4212 118 : conf: &'static PageServerConf,
4213 118 : tenant_conf: &pageserver_api::models::TenantConfig,
4214 118 : ) -> Duration {
4215 118 : tenant_conf
4216 118 : .lsn_lease_length
4217 118 : .unwrap_or(conf.default_tenant_conf.lsn_lease_length)
4218 118 : }
4219 :
4220 0 : pub fn get_timeline_offloading_enabled(&self) -> bool {
4221 0 : if self.conf.timeline_offloading {
4222 0 : return true;
4223 0 : }
4224 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
4225 0 : tenant_conf
4226 0 : .timeline_offloading
4227 0 : .unwrap_or(self.conf.default_tenant_conf.timeline_offloading)
4228 0 : }
4229 :
4230 : /// Generate an up-to-date TenantManifest based on the state of this Tenant.
4231 119 : fn build_tenant_manifest(&self) -> TenantManifest {
4232 : // Collect the offloaded timelines, and sort them for deterministic output.
4233 119 : let offloaded_timelines = self
4234 119 : .timelines_offloaded
4235 119 : .lock()
4236 119 : .unwrap()
4237 119 : .values()
4238 119 : .map(|tli| tli.manifest())
4239 119 : .sorted_by_key(|m| m.timeline_id)
4240 119 : .collect_vec();
4241 :
4242 119 : TenantManifest {
4243 119 : version: LATEST_TENANT_MANIFEST_VERSION,
4244 119 : stripe_size: Some(self.get_shard_stripe_size()),
4245 119 : offloaded_timelines,
4246 119 : }
4247 119 : }
4248 :
4249 1 : pub fn update_tenant_config<
4250 1 : F: Fn(
4251 1 : pageserver_api::models::TenantConfig,
4252 1 : ) -> anyhow::Result<pageserver_api::models::TenantConfig>,
4253 1 : >(
4254 1 : &self,
4255 1 : update: F,
4256 1 : ) -> anyhow::Result<pageserver_api::models::TenantConfig> {
4257 : // Use read-copy-update in order to avoid overwriting the location config
4258 : // state if this races with [`TenantShard::set_new_location_config`]. Note that
4259 : // this race is not possible if both request types come from the storage
4260 : // controller (as they should!) because an exclusive op lock is required
4261 : // on the storage controller side.
4262 :
4263 1 : self.tenant_conf
4264 1 : .try_rcu(|attached_conf| -> Result<_, anyhow::Error> {
4265 1 : Ok(Arc::new(AttachedTenantConf {
4266 1 : tenant_conf: update(attached_conf.tenant_conf.clone())?,
4267 1 : location: attached_conf.location,
4268 1 : lsn_lease_deadline: attached_conf.lsn_lease_deadline,
4269 : }))
4270 1 : })?;
4271 :
4272 1 : let updated = self.tenant_conf.load();
4273 :
4274 1 : self.tenant_conf_updated(&updated.tenant_conf);
4275 : // Don't hold self.timelines.lock() during the notifies.
4276 : // There's no risk of deadlock right now, but there could be if we consolidate
4277 : // mutexes in struct Timeline in the future.
4278 1 : let timelines = self.list_timelines();
4279 1 : for timeline in timelines {
4280 0 : timeline.tenant_conf_updated(&updated);
4281 0 : }
4282 :
4283 1 : Ok(updated.tenant_conf.clone())
4284 1 : }
4285 :
4286 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
4287 0 : let new_tenant_conf = new_conf.tenant_conf.clone();
4288 :
4289 0 : self.tenant_conf.store(Arc::new(new_conf.clone()));
4290 :
4291 0 : self.tenant_conf_updated(&new_tenant_conf);
4292 : // Don't hold self.timelines.lock() during the notifies.
4293 : // There's no risk of deadlock right now, but there could be if we consolidate
4294 : // mutexes in struct Timeline in the future.
4295 0 : let timelines = self.list_timelines();
4296 0 : for timeline in timelines {
4297 0 : timeline.tenant_conf_updated(&new_conf);
4298 0 : }
4299 0 : }
4300 :
4301 119 : fn get_pagestream_throttle_config(
4302 119 : psconf: &'static PageServerConf,
4303 119 : overrides: &pageserver_api::models::TenantConfig,
4304 119 : ) -> throttle::Config {
4305 119 : overrides
4306 119 : .timeline_get_throttle
4307 119 : .clone()
4308 119 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
4309 119 : }
4310 :
4311 1 : pub(crate) fn tenant_conf_updated(&self, new_conf: &pageserver_api::models::TenantConfig) {
4312 1 : let conf = Self::get_pagestream_throttle_config(self.conf, new_conf);
4313 1 : self.pagestream_throttle.reconfigure(conf)
4314 1 : }
4315 :
4316 : /// Helper function to create a new Timeline struct.
4317 : ///
4318 : /// The returned Timeline is in Loading state. The caller is responsible for
4319 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
4320 : /// map.
4321 : ///
4322 : /// `validate_ancestor == false` is used when a timeline is created for deletion
4323 : /// and we might not have the ancestor present anymore which is fine for to be
4324 : /// deleted timelines.
4325 : #[allow(clippy::too_many_arguments)]
4326 234 : fn create_timeline_struct(
4327 234 : &self,
4328 234 : new_timeline_id: TimelineId,
4329 234 : new_metadata: &TimelineMetadata,
4330 234 : previous_heatmap: Option<PreviousHeatmap>,
4331 234 : ancestor: Option<Arc<Timeline>>,
4332 234 : resources: TimelineResources,
4333 234 : cause: CreateTimelineCause,
4334 234 : create_idempotency: CreateTimelineIdempotency,
4335 234 : gc_compaction_state: Option<GcCompactionState>,
4336 234 : rel_size_v2_status: Option<RelSizeMigration>,
4337 234 : ctx: &RequestContext,
4338 234 : ) -> anyhow::Result<(Arc<Timeline>, RequestContext)> {
4339 234 : let state = match cause {
4340 : CreateTimelineCause::Load => {
4341 234 : let ancestor_id = new_metadata.ancestor_timeline();
4342 234 : anyhow::ensure!(
4343 234 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
4344 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
4345 : );
4346 234 : TimelineState::Loading
4347 : }
4348 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
4349 : };
4350 :
4351 234 : let pg_version = new_metadata.pg_version();
4352 :
4353 234 : let timeline = Timeline::new(
4354 234 : self.conf,
4355 234 : Arc::clone(&self.tenant_conf),
4356 234 : new_metadata,
4357 234 : previous_heatmap,
4358 234 : ancestor,
4359 234 : new_timeline_id,
4360 234 : self.tenant_shard_id,
4361 234 : self.generation,
4362 234 : self.shard_identity,
4363 234 : self.walredo_mgr.clone(),
4364 234 : resources,
4365 234 : pg_version,
4366 234 : state,
4367 234 : self.attach_wal_lag_cooldown.clone(),
4368 234 : create_idempotency,
4369 234 : gc_compaction_state,
4370 234 : rel_size_v2_status,
4371 234 : self.cancel.child_token(),
4372 : );
4373 :
4374 234 : let timeline_ctx = RequestContextBuilder::from(ctx)
4375 234 : .scope(context::Scope::new_timeline(&timeline))
4376 234 : .detached_child();
4377 :
4378 234 : Ok((timeline, timeline_ctx))
4379 234 : }
4380 :
4381 : /// [`TenantShard::shutdown`] must be called before dropping the returned [`TenantShard`] object
4382 : /// to ensure proper cleanup of background tasks and metrics.
4383 : //
4384 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
4385 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
4386 : #[allow(clippy::too_many_arguments)]
4387 118 : fn new(
4388 118 : state: TenantState,
4389 118 : conf: &'static PageServerConf,
4390 118 : attached_conf: AttachedTenantConf,
4391 118 : shard_identity: ShardIdentity,
4392 118 : walredo_mgr: Option<Arc<WalRedoManager>>,
4393 118 : tenant_shard_id: TenantShardId,
4394 118 : remote_storage: GenericRemoteStorage,
4395 118 : deletion_queue_client: DeletionQueueClient,
4396 118 : l0_flush_global_state: L0FlushGlobalState,
4397 118 : basebackup_cache: Arc<BasebackupCache>,
4398 118 : feature_resolver: FeatureResolver,
4399 118 : ) -> TenantShard {
4400 118 : assert!(!attached_conf.location.generation.is_none());
4401 :
4402 118 : let (state, mut rx) = watch::channel(state);
4403 :
4404 118 : tokio::spawn(async move {
4405 : // reflect tenant state in metrics:
4406 : // - global per tenant state: TENANT_STATE_METRIC
4407 : // - "set" of broken tenants: BROKEN_TENANTS_SET
4408 : //
4409 : // set of broken tenants should not have zero counts so that it remains accessible for
4410 : // alerting.
4411 :
4412 118 : let tid = tenant_shard_id.to_string();
4413 118 : let shard_id = tenant_shard_id.shard_slug().to_string();
4414 118 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
4415 :
4416 236 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
4417 236 : ([state.into()], matches!(state, TenantState::Broken { .. }))
4418 236 : }
4419 :
4420 118 : let mut tuple = inspect_state(&rx.borrow_and_update());
4421 :
4422 118 : let is_broken = tuple.1;
4423 118 : let mut counted_broken = if is_broken {
4424 : // add the id to the set right away, there should not be any updates on the channel
4425 : // after before tenant is removed, if ever
4426 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4427 0 : true
4428 : } else {
4429 118 : false
4430 : };
4431 :
4432 : loop {
4433 236 : let labels = &tuple.0;
4434 236 : let current = TENANT_STATE_METRIC.with_label_values(labels);
4435 236 : current.inc();
4436 :
4437 236 : if rx.changed().await.is_err() {
4438 : // tenant has been dropped
4439 7 : current.dec();
4440 7 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
4441 7 : break;
4442 118 : }
4443 :
4444 118 : current.dec();
4445 118 : tuple = inspect_state(&rx.borrow_and_update());
4446 :
4447 118 : let is_broken = tuple.1;
4448 118 : if is_broken && !counted_broken {
4449 0 : counted_broken = true;
4450 0 : // insert the tenant_id (back) into the set while avoiding needless counter
4451 0 : // access
4452 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
4453 118 : }
4454 : }
4455 7 : });
4456 :
4457 118 : TenantShard {
4458 118 : tenant_shard_id,
4459 118 : shard_identity,
4460 118 : generation: attached_conf.location.generation,
4461 118 : conf,
4462 118 : // using now here is good enough approximation to catch tenants with really long
4463 118 : // activation times.
4464 118 : constructed_at: Instant::now(),
4465 118 : timelines: Mutex::new(HashMap::new()),
4466 118 : timelines_creating: Mutex::new(HashSet::new()),
4467 118 : timelines_offloaded: Mutex::new(HashMap::new()),
4468 118 : timelines_importing: Mutex::new(HashMap::new()),
4469 118 : remote_tenant_manifest: Default::default(),
4470 118 : gc_cs: tokio::sync::Mutex::new(()),
4471 118 : walredo_mgr,
4472 118 : remote_storage,
4473 118 : deletion_queue_client,
4474 118 : state,
4475 118 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
4476 118 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
4477 118 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
4478 118 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
4479 118 : format!("compaction-{tenant_shard_id}"),
4480 118 : 5,
4481 118 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
4482 118 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
4483 118 : // use an extremely long backoff.
4484 118 : Some(Duration::from_secs(3600 * 24)),
4485 118 : )),
4486 118 : l0_compaction_trigger: Arc::new(Notify::new()),
4487 118 : scheduled_compaction_tasks: Mutex::new(Default::default()),
4488 118 : activate_now_sem: tokio::sync::Semaphore::new(0),
4489 118 : attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
4490 118 : cancel: CancellationToken::default(),
4491 118 : gate: Gate::default(),
4492 118 : pagestream_throttle: Arc::new(throttle::Throttle::new(
4493 118 : TenantShard::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf),
4494 118 : )),
4495 118 : pagestream_throttle_metrics: Arc::new(
4496 118 : crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id),
4497 118 : ),
4498 118 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
4499 118 : ongoing_timeline_detach: std::sync::Mutex::default(),
4500 118 : gc_block: Default::default(),
4501 118 : l0_flush_global_state,
4502 118 : basebackup_cache,
4503 118 : feature_resolver: Arc::new(TenantFeatureResolver::new(
4504 118 : feature_resolver,
4505 118 : tenant_shard_id.tenant_id,
4506 118 : )),
4507 118 : }
4508 118 : }
4509 :
4510 : /// Locate and load config
4511 0 : pub(super) fn load_tenant_config(
4512 0 : conf: &'static PageServerConf,
4513 0 : tenant_shard_id: &TenantShardId,
4514 0 : ) -> Result<LocationConf, LoadConfigError> {
4515 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4516 :
4517 0 : info!("loading tenant configuration from {config_path}");
4518 :
4519 : // load and parse file
4520 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
4521 0 : match e.kind() {
4522 : std::io::ErrorKind::NotFound => {
4523 : // The config should almost always exist for a tenant directory:
4524 : // - When attaching a tenant, the config is the first thing we write
4525 : // - When detaching a tenant, we atomically move the directory to a tmp location
4526 : // before deleting contents.
4527 : //
4528 : // The very rare edge case that can result in a missing config is if we crash during attach
4529 : // between creating directory and writing config. Callers should handle that as if the
4530 : // directory didn't exist.
4531 :
4532 0 : LoadConfigError::NotFound(config_path)
4533 : }
4534 : _ => {
4535 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
4536 : // that we cannot cleanly recover
4537 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
4538 : }
4539 : }
4540 0 : })?;
4541 :
4542 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
4543 0 : }
4544 :
4545 : /// Stores a tenant location config to disk.
4546 : ///
4547 : /// NB: make sure to call `ShardIdentity::assert_equal` before persisting a new config, to avoid
4548 : /// changes to shard parameters that may result in data corruption.
4549 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4550 : pub(super) async fn persist_tenant_config(
4551 : conf: &'static PageServerConf,
4552 : tenant_shard_id: &TenantShardId,
4553 : location_conf: &LocationConf,
4554 : ) -> std::io::Result<()> {
4555 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
4556 :
4557 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
4558 : }
4559 :
4560 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
4561 : pub(super) async fn persist_tenant_config_at(
4562 : tenant_shard_id: &TenantShardId,
4563 : config_path: &Utf8Path,
4564 : location_conf: &LocationConf,
4565 : ) -> std::io::Result<()> {
4566 : debug!("persisting tenantconf to {config_path}");
4567 :
4568 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
4569 : # It is read in case of pageserver restart.
4570 : "#
4571 : .to_string();
4572 :
4573 0 : fail::fail_point!("tenant-config-before-write", |_| {
4574 0 : Err(std::io::Error::other("tenant-config-before-write"))
4575 0 : });
4576 :
4577 : // Convert the config to a toml file.
4578 : conf_content +=
4579 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
4580 :
4581 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
4582 :
4583 : let conf_content = conf_content.into_bytes();
4584 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
4585 : }
4586 :
4587 : //
4588 : // How garbage collection works:
4589 : //
4590 : // +--bar------------->
4591 : // /
4592 : // +----+-----foo---------------->
4593 : // /
4594 : // ----main--+-------------------------->
4595 : // \
4596 : // +-----baz-------->
4597 : //
4598 : //
4599 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
4600 : // `gc_infos` are being refreshed
4601 : // 2. Scan collected timelines, and on each timeline, make note of the
4602 : // all the points where other timelines have been branched off.
4603 : // We will refrain from removing page versions at those LSNs.
4604 : // 3. For each timeline, scan all layer files on the timeline.
4605 : // Remove all files for which a newer file exists and which
4606 : // don't cover any branch point LSNs.
4607 : //
4608 : // TODO:
4609 : // - if a relation has a non-incremental persistent layer on a child branch, then we
4610 : // don't need to keep that in the parent anymore. But currently
4611 : // we do.
4612 377 : async fn gc_iteration_internal(
4613 377 : &self,
4614 377 : target_timeline_id: Option<TimelineId>,
4615 377 : horizon: u64,
4616 377 : pitr: Duration,
4617 377 : cancel: &CancellationToken,
4618 377 : ctx: &RequestContext,
4619 377 : ) -> Result<GcResult, GcError> {
4620 377 : let mut totals: GcResult = Default::default();
4621 377 : let now = Instant::now();
4622 :
4623 377 : let gc_timelines = self
4624 377 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4625 377 : .await?;
4626 :
4627 377 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
4628 :
4629 : // If there is nothing to GC, we don't want any messages in the INFO log.
4630 377 : if !gc_timelines.is_empty() {
4631 377 : info!("{} timelines need GC", gc_timelines.len());
4632 : } else {
4633 0 : debug!("{} timelines need GC", gc_timelines.len());
4634 : }
4635 :
4636 : // Perform GC for each timeline.
4637 : //
4638 : // Note that we don't hold the `TenantShard::gc_cs` lock here because we don't want to delay the
4639 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
4640 : // with branch creation.
4641 : //
4642 : // See comments in [`TenantShard::branch_timeline`] for more information about why branch
4643 : // creation task can run concurrently with timeline's GC iteration.
4644 754 : for timeline in gc_timelines {
4645 377 : if cancel.is_cancelled() {
4646 : // We were requested to shut down. Stop and return with the progress we
4647 : // made.
4648 0 : break;
4649 377 : }
4650 377 : let result = match timeline.gc().await {
4651 : Err(GcError::TimelineCancelled) => {
4652 0 : if target_timeline_id.is_some() {
4653 : // If we were targetting this specific timeline, surface cancellation to caller
4654 0 : return Err(GcError::TimelineCancelled);
4655 : } else {
4656 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
4657 : // skip past this and proceed to try GC on other timelines.
4658 0 : continue;
4659 : }
4660 : }
4661 377 : r => r?,
4662 : };
4663 377 : totals += result;
4664 : }
4665 :
4666 377 : totals.elapsed = now.elapsed();
4667 377 : Ok(totals)
4668 377 : }
4669 :
4670 : /// Refreshes the Timeline::gc_info for all timelines, returning the
4671 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
4672 : /// [`TenantShard::get_gc_horizon`].
4673 : ///
4674 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
4675 2 : pub(crate) async fn refresh_gc_info(
4676 2 : &self,
4677 2 : cancel: &CancellationToken,
4678 2 : ctx: &RequestContext,
4679 2 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4680 : // since this method can now be called at different rates than the configured gc loop, it
4681 : // might be that these configuration values get applied faster than what it was previously,
4682 : // since these were only read from the gc task.
4683 2 : let horizon = self.get_gc_horizon();
4684 2 : let pitr = self.get_pitr_interval();
4685 :
4686 : // refresh all timelines
4687 2 : let target_timeline_id = None;
4688 :
4689 2 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
4690 2 : .await
4691 2 : }
4692 :
4693 : /// Populate all Timelines' `GcInfo` with information about their children. We do not set the
4694 : /// PITR cutoffs here, because that requires I/O: this is done later, before GC, by [`Self::refresh_gc_info_internal`]
4695 : ///
4696 : /// Subsequently, parent-child relationships are updated incrementally inside [`Timeline::new`] and [`Timeline::drop`].
4697 118 : fn initialize_gc_info(
4698 118 : &self,
4699 118 : timelines: &std::sync::MutexGuard<HashMap<TimelineId, Arc<Timeline>>>,
4700 118 : timelines_offloaded: &std::sync::MutexGuard<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
4701 118 : restrict_to_timeline: Option<TimelineId>,
4702 118 : ) {
4703 118 : if restrict_to_timeline.is_none() {
4704 : // This function must be called before activation: after activation timeline create/delete operations
4705 : // might happen, and this function is not safe to run concurrently with those.
4706 118 : assert!(!self.is_active());
4707 0 : }
4708 :
4709 : // Scan all timelines. For each timeline, remember the timeline ID and
4710 : // the branch point where it was created.
4711 118 : let mut all_branchpoints: BTreeMap<TimelineId, Vec<(Lsn, TimelineId, MaybeOffloaded)>> =
4712 118 : BTreeMap::new();
4713 118 : timelines.iter().for_each(|(timeline_id, timeline_entry)| {
4714 3 : if let Some(ancestor_timeline_id) = &timeline_entry.get_ancestor_timeline_id() {
4715 1 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4716 1 : ancestor_children.push((
4717 1 : timeline_entry.get_ancestor_lsn(),
4718 1 : *timeline_id,
4719 1 : MaybeOffloaded::No,
4720 1 : ));
4721 2 : }
4722 3 : });
4723 118 : timelines_offloaded
4724 118 : .iter()
4725 118 : .for_each(|(timeline_id, timeline_entry)| {
4726 0 : let Some(ancestor_timeline_id) = &timeline_entry.ancestor_timeline_id else {
4727 0 : return;
4728 : };
4729 0 : let Some(retain_lsn) = timeline_entry.ancestor_retain_lsn else {
4730 0 : return;
4731 : };
4732 0 : let ancestor_children = all_branchpoints.entry(*ancestor_timeline_id).or_default();
4733 0 : ancestor_children.push((retain_lsn, *timeline_id, MaybeOffloaded::Yes));
4734 0 : });
4735 :
4736 : // The number of bytes we always keep, irrespective of PITR: this is a constant across timelines
4737 118 : let horizon = self.get_gc_horizon();
4738 :
4739 : // Populate each timeline's GcInfo with information about its child branches
4740 118 : let timelines_to_write = if let Some(timeline_id) = restrict_to_timeline {
4741 0 : itertools::Either::Left(timelines.get(&timeline_id).into_iter())
4742 : } else {
4743 118 : itertools::Either::Right(timelines.values())
4744 : };
4745 121 : for timeline in timelines_to_write {
4746 3 : let mut branchpoints: Vec<(Lsn, TimelineId, MaybeOffloaded)> = all_branchpoints
4747 3 : .remove(&timeline.timeline_id)
4748 3 : .unwrap_or_default();
4749 :
4750 3 : branchpoints.sort_by_key(|b| b.0);
4751 :
4752 3 : let mut target = timeline.gc_info.write().unwrap();
4753 :
4754 3 : target.retain_lsns = branchpoints;
4755 :
4756 3 : let space_cutoff = timeline
4757 3 : .get_last_record_lsn()
4758 3 : .checked_sub(horizon)
4759 3 : .unwrap_or(Lsn(0));
4760 :
4761 3 : target.cutoffs = GcCutoffs {
4762 3 : space: space_cutoff,
4763 3 : time: None,
4764 3 : };
4765 : }
4766 118 : }
4767 :
4768 379 : async fn refresh_gc_info_internal(
4769 379 : &self,
4770 379 : target_timeline_id: Option<TimelineId>,
4771 379 : horizon: u64,
4772 379 : pitr: Duration,
4773 379 : cancel: &CancellationToken,
4774 379 : ctx: &RequestContext,
4775 379 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
4776 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
4777 : // currently visible timelines.
4778 379 : let timelines = self
4779 379 : .timelines
4780 379 : .lock()
4781 379 : .unwrap()
4782 379 : .values()
4783 1663 : .filter(|tl| match target_timeline_id.as_ref() {
4784 1655 : Some(target) => &tl.timeline_id == target,
4785 8 : None => true,
4786 1663 : })
4787 379 : .cloned()
4788 379 : .collect::<Vec<_>>();
4789 :
4790 379 : if target_timeline_id.is_some() && timelines.is_empty() {
4791 : // We were to act on a particular timeline and it wasn't found
4792 0 : return Err(GcError::TimelineNotFound);
4793 379 : }
4794 :
4795 379 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
4796 379 : HashMap::with_capacity(timelines.len());
4797 :
4798 : // Ensures all timelines use the same start time when computing the time cutoff.
4799 379 : let now_ts_for_pitr_calc = SystemTime::now();
4800 385 : for timeline in timelines.iter() {
4801 385 : let ctx = &ctx.with_scope_timeline(timeline);
4802 385 : let cutoff = timeline
4803 385 : .get_last_record_lsn()
4804 385 : .checked_sub(horizon)
4805 385 : .unwrap_or(Lsn(0));
4806 :
4807 385 : let cutoffs = timeline
4808 385 : .find_gc_cutoffs(now_ts_for_pitr_calc, cutoff, pitr, cancel, ctx)
4809 385 : .await?;
4810 385 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
4811 385 : assert!(old.is_none());
4812 : }
4813 :
4814 379 : if !self.is_active() || self.cancel.is_cancelled() {
4815 0 : return Err(GcError::TenantCancelled);
4816 379 : }
4817 :
4818 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
4819 : // because that will stall branch creation.
4820 379 : let gc_cs = self.gc_cs.lock().await;
4821 :
4822 : // Ok, we now know all the branch points.
4823 : // Update the GC information for each timeline.
4824 379 : let mut gc_timelines = Vec::with_capacity(timelines.len());
4825 764 : for timeline in timelines {
4826 : // We filtered the timeline list above
4827 385 : if let Some(target_timeline_id) = target_timeline_id {
4828 377 : assert_eq!(target_timeline_id, timeline.timeline_id);
4829 8 : }
4830 :
4831 : {
4832 385 : let mut target = timeline.gc_info.write().unwrap();
4833 :
4834 : // Cull any expired leases
4835 385 : let now = SystemTime::now();
4836 385 : target.leases.retain(|_, lease| !lease.is_expired(&now));
4837 :
4838 385 : timeline
4839 385 : .metrics
4840 385 : .valid_lsn_lease_count_gauge
4841 385 : .set(target.leases.len() as u64);
4842 :
4843 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
4844 385 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
4845 56 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
4846 6 : target.within_ancestor_pitr =
4847 6 : Some(timeline.get_ancestor_lsn()) >= ancestor_gc_cutoffs.time;
4848 50 : }
4849 329 : }
4850 :
4851 : // Update metrics that depend on GC state
4852 385 : timeline
4853 385 : .metrics
4854 385 : .archival_size
4855 385 : .set(if target.within_ancestor_pitr {
4856 0 : timeline.metrics.current_logical_size_gauge.get()
4857 : } else {
4858 385 : 0
4859 : });
4860 385 : if let Some(time_cutoff) = target.cutoffs.time {
4861 319 : timeline.metrics.pitr_history_size.set(
4862 319 : timeline
4863 319 : .get_last_record_lsn()
4864 319 : .checked_sub(time_cutoff)
4865 319 : .unwrap_or_default()
4866 319 : .0,
4867 319 : );
4868 319 : }
4869 :
4870 : // Apply the cutoffs we found to the Timeline's GcInfo. Why might we _not_ have cutoffs for a timeline?
4871 : // - this timeline was created while we were finding cutoffs
4872 : // - lsn for timestamp search fails for this timeline repeatedly
4873 385 : if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
4874 385 : let original_cutoffs = target.cutoffs.clone();
4875 : // GC cutoffs should never go back
4876 385 : target.cutoffs = GcCutoffs {
4877 385 : space: cutoffs.space.max(original_cutoffs.space),
4878 385 : time: cutoffs.time.max(original_cutoffs.time),
4879 385 : }
4880 0 : }
4881 : }
4882 :
4883 385 : gc_timelines.push(timeline);
4884 : }
4885 379 : drop(gc_cs);
4886 379 : Ok(gc_timelines)
4887 379 : }
4888 :
4889 : /// A substitute for `branch_timeline` for use in unit tests.
4890 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
4891 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
4892 : /// timeline background tasks are launched, except the flush loop.
4893 : #[cfg(test)]
4894 119 : async fn branch_timeline_test(
4895 119 : self: &Arc<Self>,
4896 119 : src_timeline: &Arc<Timeline>,
4897 119 : dst_id: TimelineId,
4898 119 : ancestor_lsn: Option<Lsn>,
4899 119 : ctx: &RequestContext,
4900 119 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
4901 119 : let tl = self
4902 119 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, ctx)
4903 119 : .await?
4904 117 : .into_timeline_for_test();
4905 117 : tl.set_state(TimelineState::Active);
4906 117 : Ok(tl)
4907 119 : }
4908 :
4909 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
4910 : #[cfg(test)]
4911 : #[allow(clippy::too_many_arguments)]
4912 6 : pub async fn branch_timeline_test_with_layers(
4913 6 : self: &Arc<Self>,
4914 6 : src_timeline: &Arc<Timeline>,
4915 6 : dst_id: TimelineId,
4916 6 : ancestor_lsn: Option<Lsn>,
4917 6 : ctx: &RequestContext,
4918 6 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
4919 6 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
4920 6 : end_lsn: Lsn,
4921 6 : ) -> anyhow::Result<Arc<Timeline>> {
4922 : use checks::check_valid_layermap;
4923 : use itertools::Itertools;
4924 :
4925 6 : let tline = self
4926 6 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
4927 6 : .await?;
4928 6 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
4929 6 : ancestor_lsn
4930 : } else {
4931 0 : tline.get_last_record_lsn()
4932 : };
4933 6 : assert!(end_lsn >= ancestor_lsn);
4934 6 : tline.force_advance_lsn(end_lsn);
4935 9 : for deltas in delta_layer_desc {
4936 3 : tline
4937 3 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
4938 3 : .await?;
4939 : }
4940 8 : for (lsn, images) in image_layer_desc {
4941 2 : tline
4942 2 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
4943 2 : .await?;
4944 : }
4945 6 : let layer_names = tline
4946 6 : .layers
4947 6 : .read(LayerManagerLockHolder::Testing)
4948 6 : .await
4949 6 : .layer_map()
4950 6 : .unwrap()
4951 6 : .iter_historic_layers()
4952 6 : .map(|layer| layer.layer_name())
4953 6 : .collect_vec();
4954 6 : if let Some(err) = check_valid_layermap(&layer_names) {
4955 0 : bail!("invalid layermap: {err}");
4956 6 : }
4957 6 : Ok(tline)
4958 6 : }
4959 :
4960 : /// Branch an existing timeline.
4961 0 : async fn branch_timeline(
4962 0 : self: &Arc<Self>,
4963 0 : src_timeline: &Arc<Timeline>,
4964 0 : dst_id: TimelineId,
4965 0 : start_lsn: Option<Lsn>,
4966 0 : ctx: &RequestContext,
4967 0 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4968 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
4969 0 : .await
4970 0 : }
4971 :
4972 119 : async fn branch_timeline_impl(
4973 119 : self: &Arc<Self>,
4974 119 : src_timeline: &Arc<Timeline>,
4975 119 : dst_id: TimelineId,
4976 119 : start_lsn: Option<Lsn>,
4977 119 : ctx: &RequestContext,
4978 119 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
4979 119 : let src_id = src_timeline.timeline_id;
4980 :
4981 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
4982 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
4983 : // valid while we are creating the branch.
4984 119 : let _gc_cs = self.gc_cs.lock().await;
4985 :
4986 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
4987 119 : let start_lsn = start_lsn.unwrap_or_else(|| {
4988 1 : let lsn = src_timeline.get_last_record_lsn();
4989 1 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
4990 1 : lsn
4991 1 : });
4992 :
4993 : // we finally have determined the ancestor_start_lsn, so we can get claim exclusivity now
4994 119 : let timeline_create_guard = match self
4995 119 : .start_creating_timeline(
4996 119 : dst_id,
4997 119 : CreateTimelineIdempotency::Branch {
4998 119 : ancestor_timeline_id: src_timeline.timeline_id,
4999 119 : ancestor_start_lsn: start_lsn,
5000 119 : },
5001 119 : )
5002 119 : .await?
5003 : {
5004 119 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5005 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5006 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5007 : }
5008 : };
5009 :
5010 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
5011 : // horizon on the source timeline
5012 : //
5013 : // We check it against both the planned GC cutoff stored in 'gc_info',
5014 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
5015 : // planned GC cutoff in 'gc_info' is normally larger than
5016 : // 'applied_gc_cutoff_lsn', but beware of corner cases like if you just
5017 : // changed the GC settings for the tenant to make the PITR window
5018 : // larger, but some of the data was already removed by an earlier GC
5019 : // iteration.
5020 :
5021 : // check against last actual 'latest_gc_cutoff' first
5022 119 : let applied_gc_cutoff_lsn = src_timeline.get_applied_gc_cutoff_lsn();
5023 : {
5024 119 : let gc_info = src_timeline.gc_info.read().unwrap();
5025 119 : let planned_cutoff = gc_info.min_cutoff();
5026 119 : if gc_info.lsn_covered_by_lease(start_lsn) {
5027 0 : tracing::info!(
5028 0 : "skipping comparison of {start_lsn} with gc cutoff {} and planned gc cutoff {planned_cutoff} due to lsn lease",
5029 0 : *applied_gc_cutoff_lsn
5030 : );
5031 : } else {
5032 119 : src_timeline
5033 119 : .check_lsn_is_in_scope(start_lsn, &applied_gc_cutoff_lsn)
5034 119 : .context(format!(
5035 119 : "invalid branch start lsn: less than latest GC cutoff {}",
5036 119 : *applied_gc_cutoff_lsn,
5037 : ))
5038 119 : .map_err(CreateTimelineError::AncestorLsn)?;
5039 :
5040 : // and then the planned GC cutoff
5041 117 : if start_lsn < planned_cutoff {
5042 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
5043 0 : "invalid branch start lsn: less than planned GC cutoff {planned_cutoff}"
5044 0 : )));
5045 117 : }
5046 : }
5047 : }
5048 :
5049 : //
5050 : // The branch point is valid, and we are still holding the 'gc_cs' lock
5051 : // so that GC cannot advance the GC cutoff until we are finished.
5052 : // Proceed with the branch creation.
5053 : //
5054 :
5055 : // Determine prev-LSN for the new timeline. We can only determine it if
5056 : // the timeline was branched at the current end of the source timeline.
5057 : let RecordLsn {
5058 117 : last: src_last,
5059 117 : prev: src_prev,
5060 117 : } = src_timeline.get_last_record_rlsn();
5061 117 : let dst_prev = if src_last == start_lsn {
5062 108 : Some(src_prev)
5063 : } else {
5064 9 : None
5065 : };
5066 :
5067 : // Create the metadata file, noting the ancestor of the new timeline.
5068 : // There is initially no data in it, but all the read-calls know to look
5069 : // into the ancestor.
5070 117 : let metadata = TimelineMetadata::new(
5071 117 : start_lsn,
5072 117 : dst_prev,
5073 117 : Some(src_id),
5074 117 : start_lsn,
5075 117 : *src_timeline.applied_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
5076 117 : src_timeline.initdb_lsn,
5077 117 : src_timeline.pg_version,
5078 : );
5079 :
5080 117 : let (uninitialized_timeline, _timeline_ctx) = self
5081 117 : .prepare_new_timeline(
5082 117 : dst_id,
5083 117 : &metadata,
5084 117 : timeline_create_guard,
5085 117 : start_lsn + 1,
5086 117 : Some(Arc::clone(src_timeline)),
5087 117 : Some(src_timeline.get_rel_size_v2_status()),
5088 117 : ctx,
5089 117 : )
5090 117 : .await?;
5091 :
5092 117 : let new_timeline = uninitialized_timeline.finish_creation().await?;
5093 :
5094 : // Root timeline gets its layers during creation and uploads them along with the metadata.
5095 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
5096 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
5097 : // could get incorrect information and remove more layers, than needed.
5098 : // See also https://github.com/neondatabase/neon/issues/3865
5099 117 : new_timeline
5100 117 : .remote_client
5101 117 : .schedule_index_upload_for_full_metadata_update(&metadata)
5102 117 : .context("branch initial metadata upload")?;
5103 :
5104 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5105 :
5106 117 : Ok(CreateTimelineResult::Created(new_timeline))
5107 119 : }
5108 :
5109 : /// For unit tests, make this visible so that other modules can directly create timelines
5110 : #[cfg(test)]
5111 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
5112 : pub(crate) async fn bootstrap_timeline_test(
5113 : self: &Arc<Self>,
5114 : timeline_id: TimelineId,
5115 : pg_version: PgMajorVersion,
5116 : load_existing_initdb: Option<TimelineId>,
5117 : ctx: &RequestContext,
5118 : ) -> anyhow::Result<Arc<Timeline>> {
5119 : self.bootstrap_timeline(timeline_id, pg_version, load_existing_initdb, ctx)
5120 : .await
5121 : .map_err(anyhow::Error::new)
5122 1 : .map(|r| r.into_timeline_for_test())
5123 : }
5124 :
5125 : /// Get exclusive access to the timeline ID for creation.
5126 : ///
5127 : /// Timeline-creating code paths must use this function before making changes
5128 : /// to in-memory or persistent state.
5129 : ///
5130 : /// The `state` parameter is a description of the timeline creation operation
5131 : /// we intend to perform.
5132 : /// If the timeline was already created in the meantime, we check whether this
5133 : /// request conflicts or is idempotent , based on `state`.
5134 234 : async fn start_creating_timeline(
5135 234 : self: &Arc<Self>,
5136 234 : new_timeline_id: TimelineId,
5137 234 : idempotency: CreateTimelineIdempotency,
5138 234 : ) -> Result<StartCreatingTimelineResult, CreateTimelineError> {
5139 234 : let allow_offloaded = false;
5140 234 : match self.create_timeline_create_guard(new_timeline_id, idempotency, allow_offloaded) {
5141 233 : Ok(create_guard) => {
5142 233 : pausable_failpoint!("timeline-creation-after-uninit");
5143 233 : Ok(StartCreatingTimelineResult::CreateGuard(create_guard))
5144 : }
5145 0 : Err(TimelineExclusionError::ShuttingDown) => Err(CreateTimelineError::ShuttingDown),
5146 : Err(TimelineExclusionError::AlreadyCreating) => {
5147 : // Creation is in progress, we cannot create it again, and we cannot
5148 : // check if this request matches the existing one, so caller must try
5149 : // again later.
5150 0 : Err(CreateTimelineError::AlreadyCreating)
5151 : }
5152 0 : Err(TimelineExclusionError::Other(e)) => Err(CreateTimelineError::Other(e)),
5153 : Err(TimelineExclusionError::AlreadyExists {
5154 0 : existing: TimelineOrOffloaded::Offloaded(_existing),
5155 : ..
5156 : }) => {
5157 0 : info!("timeline already exists but is offloaded");
5158 0 : Err(CreateTimelineError::Conflict)
5159 : }
5160 : Err(TimelineExclusionError::AlreadyExists {
5161 0 : existing: TimelineOrOffloaded::Importing(_existing),
5162 : ..
5163 : }) => {
5164 : // If there's a timeline already importing, then we would hit
5165 : // the [`TimelineExclusionError::AlreadyCreating`] branch above.
5166 0 : unreachable!("Importing timelines hold the creation guard")
5167 : }
5168 : Err(TimelineExclusionError::AlreadyExists {
5169 1 : existing: TimelineOrOffloaded::Timeline(existing),
5170 1 : arg,
5171 : }) => {
5172 : {
5173 1 : let existing = &existing.create_idempotency;
5174 1 : let _span = info_span!("idempotency_check", ?existing, ?arg).entered();
5175 1 : debug!("timeline already exists");
5176 :
5177 1 : match (existing, &arg) {
5178 : // FailWithConflict => no idempotency check
5179 : (CreateTimelineIdempotency::FailWithConflict, _)
5180 : | (_, CreateTimelineIdempotency::FailWithConflict) => {
5181 1 : warn!("timeline already exists, failing request");
5182 1 : return Err(CreateTimelineError::Conflict);
5183 : }
5184 : // Idempotent <=> CreateTimelineIdempotency is identical
5185 0 : (x, y) if x == y => {
5186 0 : info!(
5187 0 : "timeline already exists and idempotency matches, succeeding request"
5188 : );
5189 : // fallthrough
5190 : }
5191 : (_, _) => {
5192 0 : warn!("idempotency conflict, failing request");
5193 0 : return Err(CreateTimelineError::Conflict);
5194 : }
5195 : }
5196 : }
5197 :
5198 0 : Ok(StartCreatingTimelineResult::Idempotent(existing))
5199 : }
5200 : }
5201 234 : }
5202 :
5203 0 : async fn upload_initdb(
5204 0 : &self,
5205 0 : timelines_path: &Utf8PathBuf,
5206 0 : pgdata_path: &Utf8PathBuf,
5207 0 : timeline_id: &TimelineId,
5208 0 : ) -> anyhow::Result<()> {
5209 0 : let temp_path = timelines_path.join(format!(
5210 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
5211 0 : ));
5212 :
5213 0 : scopeguard::defer! {
5214 : if let Err(e) = fs::remove_file(&temp_path) {
5215 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
5216 : }
5217 : }
5218 :
5219 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
5220 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
5221 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
5222 0 : warn!(
5223 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
5224 : );
5225 0 : }
5226 :
5227 0 : pausable_failpoint!("before-initdb-upload");
5228 :
5229 0 : backoff::retry(
5230 0 : || async {
5231 0 : self::remote_timeline_client::upload_initdb_dir(
5232 0 : &self.remote_storage,
5233 0 : &self.tenant_shard_id.tenant_id,
5234 0 : timeline_id,
5235 0 : pgdata_zstd.try_clone().await?,
5236 0 : tar_zst_size,
5237 0 : &self.cancel,
5238 : )
5239 0 : .await
5240 0 : },
5241 : |_| false,
5242 : 3,
5243 : u32::MAX,
5244 0 : "persist_initdb_tar_zst",
5245 0 : &self.cancel,
5246 : )
5247 0 : .await
5248 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
5249 0 : .and_then(|x| x)
5250 0 : }
5251 :
5252 : /// - run initdb to init temporary instance and get bootstrap data
5253 : /// - after initialization completes, tar up the temp dir and upload it to S3.
5254 1 : async fn bootstrap_timeline(
5255 1 : self: &Arc<Self>,
5256 1 : timeline_id: TimelineId,
5257 1 : pg_version: PgMajorVersion,
5258 1 : load_existing_initdb: Option<TimelineId>,
5259 1 : ctx: &RequestContext,
5260 1 : ) -> Result<CreateTimelineResult, CreateTimelineError> {
5261 1 : let timeline_create_guard = match self
5262 1 : .start_creating_timeline(
5263 1 : timeline_id,
5264 1 : CreateTimelineIdempotency::Bootstrap { pg_version },
5265 1 : )
5266 1 : .await?
5267 : {
5268 1 : StartCreatingTimelineResult::CreateGuard(guard) => guard,
5269 0 : StartCreatingTimelineResult::Idempotent(timeline) => {
5270 0 : return Ok(CreateTimelineResult::Idempotent(timeline));
5271 : }
5272 : };
5273 :
5274 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
5275 : // temporary directory for basebackup files for the given timeline.
5276 :
5277 1 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
5278 1 : let pgdata_path = path_with_suffix_extension(
5279 1 : timelines_path.join(format!("basebackup-{timeline_id}")),
5280 1 : TEMP_FILE_SUFFIX,
5281 : );
5282 :
5283 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
5284 : // we won't race with other creations or existent timelines with the same path.
5285 1 : if pgdata_path.exists() {
5286 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
5287 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
5288 0 : })?;
5289 0 : tracing::info!("removed previous attempt's temporary initdb directory '{pgdata_path}'");
5290 1 : }
5291 :
5292 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
5293 1 : let pgdata_path_deferred = pgdata_path.clone();
5294 1 : scopeguard::defer! {
5295 : if let Err(e) = fs::remove_dir_all(&pgdata_path_deferred).or_else(fs_ext::ignore_not_found) {
5296 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
5297 : error!("Failed to remove temporary initdb directory '{pgdata_path_deferred}': {e}");
5298 : } else {
5299 : tracing::info!("removed temporary initdb directory '{pgdata_path_deferred}'");
5300 : }
5301 : }
5302 1 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
5303 1 : if existing_initdb_timeline_id != timeline_id {
5304 0 : let source_path = &remote_initdb_archive_path(
5305 0 : &self.tenant_shard_id.tenant_id,
5306 0 : &existing_initdb_timeline_id,
5307 0 : );
5308 0 : let dest_path =
5309 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
5310 :
5311 : // if this fails, it will get retried by retried control plane requests
5312 0 : self.remote_storage
5313 0 : .copy_object(source_path, dest_path, &self.cancel)
5314 0 : .await
5315 0 : .context("copy initdb tar")?;
5316 1 : }
5317 1 : let (initdb_tar_zst_path, initdb_tar_zst) =
5318 1 : self::remote_timeline_client::download_initdb_tar_zst(
5319 1 : self.conf,
5320 1 : &self.remote_storage,
5321 1 : &self.tenant_shard_id,
5322 1 : &existing_initdb_timeline_id,
5323 1 : &self.cancel,
5324 1 : )
5325 1 : .await
5326 1 : .context("download initdb tar")?;
5327 :
5328 1 : scopeguard::defer! {
5329 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
5330 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
5331 : }
5332 : }
5333 :
5334 1 : let buf_read =
5335 1 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
5336 1 : extract_zst_tarball(&pgdata_path, buf_read)
5337 1 : .await
5338 1 : .context("extract initdb tar")?;
5339 : } else {
5340 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
5341 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel)
5342 0 : .await
5343 0 : .context("run initdb")?;
5344 :
5345 : // Upload the created data dir to S3
5346 0 : if self.tenant_shard_id().is_shard_zero() {
5347 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
5348 0 : .await?;
5349 0 : }
5350 : }
5351 1 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
5352 :
5353 : // Import the contents of the data directory at the initial checkpoint
5354 : // LSN, and any WAL after that.
5355 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
5356 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
5357 1 : let new_metadata = TimelineMetadata::new(
5358 1 : Lsn(0),
5359 1 : None,
5360 1 : None,
5361 1 : Lsn(0),
5362 1 : pgdata_lsn,
5363 1 : pgdata_lsn,
5364 1 : pg_version,
5365 : );
5366 1 : let (mut raw_timeline, timeline_ctx) = self
5367 1 : .prepare_new_timeline(
5368 1 : timeline_id,
5369 1 : &new_metadata,
5370 1 : timeline_create_guard,
5371 1 : pgdata_lsn,
5372 1 : None,
5373 1 : None,
5374 1 : ctx,
5375 1 : )
5376 1 : .await?;
5377 :
5378 1 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
5379 1 : raw_timeline
5380 1 : .write(|unfinished_timeline| async move {
5381 1 : import_datadir::import_timeline_from_postgres_datadir(
5382 1 : &unfinished_timeline,
5383 1 : &pgdata_path,
5384 1 : pgdata_lsn,
5385 1 : &timeline_ctx,
5386 1 : )
5387 1 : .await
5388 1 : .with_context(|| {
5389 0 : format!(
5390 0 : "Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}"
5391 : )
5392 0 : })?;
5393 :
5394 1 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
5395 0 : Err(CreateTimelineError::Other(anyhow::anyhow!(
5396 0 : "failpoint before-checkpoint-new-timeline"
5397 0 : )))
5398 0 : });
5399 :
5400 1 : Ok(())
5401 2 : })
5402 1 : .await?;
5403 :
5404 : // All done!
5405 1 : let timeline = raw_timeline.finish_creation().await?;
5406 :
5407 : // Callers are responsible to wait for uploads to complete and for activating the timeline.
5408 :
5409 1 : Ok(CreateTimelineResult::Created(timeline))
5410 1 : }
5411 :
5412 231 : fn build_timeline_remote_client(&self, timeline_id: TimelineId) -> RemoteTimelineClient {
5413 231 : RemoteTimelineClient::new(
5414 231 : self.remote_storage.clone(),
5415 231 : self.deletion_queue_client.clone(),
5416 231 : self.conf,
5417 231 : self.tenant_shard_id,
5418 231 : timeline_id,
5419 231 : self.generation,
5420 231 : &self.tenant_conf.load().location,
5421 : )
5422 231 : }
5423 :
5424 : /// Builds required resources for a new timeline.
5425 231 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
5426 231 : let remote_client = self.build_timeline_remote_client(timeline_id);
5427 231 : self.get_timeline_resources_for(remote_client)
5428 231 : }
5429 :
5430 : /// Builds timeline resources for the given remote client.
5431 234 : fn get_timeline_resources_for(&self, remote_client: RemoteTimelineClient) -> TimelineResources {
5432 234 : TimelineResources {
5433 234 : remote_client,
5434 234 : pagestream_throttle: self.pagestream_throttle.clone(),
5435 234 : pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
5436 234 : l0_compaction_trigger: self.l0_compaction_trigger.clone(),
5437 234 : l0_flush_global_state: self.l0_flush_global_state.clone(),
5438 234 : basebackup_cache: self.basebackup_cache.clone(),
5439 234 : feature_resolver: self.feature_resolver.clone(),
5440 234 : }
5441 234 : }
5442 :
5443 : /// Creates intermediate timeline structure and its files.
5444 : ///
5445 : /// An empty layer map is initialized, and new data and WAL can be imported starting
5446 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
5447 : /// `finish_creation` to insert the Timeline into the timelines map.
5448 : #[allow(clippy::too_many_arguments)]
5449 231 : async fn prepare_new_timeline<'a>(
5450 231 : &'a self,
5451 231 : new_timeline_id: TimelineId,
5452 231 : new_metadata: &TimelineMetadata,
5453 231 : create_guard: TimelineCreateGuard,
5454 231 : start_lsn: Lsn,
5455 231 : ancestor: Option<Arc<Timeline>>,
5456 231 : rel_size_v2_status: Option<RelSizeMigration>,
5457 231 : ctx: &RequestContext,
5458 231 : ) -> anyhow::Result<(UninitializedTimeline<'a>, RequestContext)> {
5459 231 : let tenant_shard_id = self.tenant_shard_id;
5460 :
5461 231 : let resources = self.build_timeline_resources(new_timeline_id);
5462 231 : resources
5463 231 : .remote_client
5464 231 : .init_upload_queue_for_empty_remote(new_metadata, rel_size_v2_status.clone())?;
5465 :
5466 231 : let (timeline_struct, timeline_ctx) = self
5467 231 : .create_timeline_struct(
5468 231 : new_timeline_id,
5469 231 : new_metadata,
5470 231 : None,
5471 231 : ancestor,
5472 231 : resources,
5473 231 : CreateTimelineCause::Load,
5474 231 : create_guard.idempotency.clone(),
5475 231 : None,
5476 231 : rel_size_v2_status,
5477 231 : ctx,
5478 : )
5479 231 : .context("Failed to create timeline data structure")?;
5480 :
5481 231 : timeline_struct.init_empty_layer_map(start_lsn);
5482 :
5483 231 : if let Err(e) = self
5484 231 : .create_timeline_files(&create_guard.timeline_path)
5485 231 : .await
5486 : {
5487 0 : error!(
5488 0 : "Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}"
5489 : );
5490 0 : cleanup_timeline_directory(create_guard);
5491 0 : return Err(e);
5492 231 : }
5493 :
5494 231 : debug!(
5495 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
5496 : );
5497 :
5498 231 : Ok((
5499 231 : UninitializedTimeline::new(
5500 231 : self,
5501 231 : new_timeline_id,
5502 231 : Some((timeline_struct, create_guard)),
5503 231 : ),
5504 231 : timeline_ctx,
5505 231 : ))
5506 231 : }
5507 :
5508 231 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
5509 231 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
5510 :
5511 231 : fail::fail_point!("after-timeline-dir-creation", |_| {
5512 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
5513 0 : });
5514 :
5515 231 : Ok(())
5516 231 : }
5517 :
5518 : /// Get a guard that provides exclusive access to the timeline directory, preventing
5519 : /// concurrent attempts to create the same timeline.
5520 : ///
5521 : /// The `allow_offloaded` parameter controls whether to tolerate the existence of
5522 : /// offloaded timelines or not.
5523 234 : fn create_timeline_create_guard(
5524 234 : self: &Arc<Self>,
5525 234 : timeline_id: TimelineId,
5526 234 : idempotency: CreateTimelineIdempotency,
5527 234 : allow_offloaded: bool,
5528 234 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
5529 234 : let tenant_shard_id = self.tenant_shard_id;
5530 :
5531 234 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
5532 :
5533 234 : let create_guard = TimelineCreateGuard::new(
5534 234 : self,
5535 234 : timeline_id,
5536 234 : timeline_path.clone(),
5537 234 : idempotency,
5538 234 : allow_offloaded,
5539 1 : )?;
5540 :
5541 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
5542 : // for creation.
5543 : // A timeline directory should never exist on disk already:
5544 : // - a previous failed creation would have cleaned up after itself
5545 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
5546 : //
5547 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
5548 : // this error may indicate a bug in cleanup on failed creations.
5549 233 : if timeline_path.exists() {
5550 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
5551 0 : "Timeline directory already exists! This is a bug."
5552 0 : )));
5553 233 : }
5554 :
5555 233 : Ok(create_guard)
5556 234 : }
5557 :
5558 : /// Gathers inputs from all of the timelines to produce a sizing model input.
5559 : ///
5560 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
5561 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5562 : pub async fn gather_size_inputs(
5563 : &self,
5564 : // `max_retention_period` overrides the cutoff that is used to calculate the size
5565 : // (only if it is shorter than the real cutoff).
5566 : max_retention_period: Option<u64>,
5567 : cause: LogicalSizeCalculationCause,
5568 : cancel: &CancellationToken,
5569 : ctx: &RequestContext,
5570 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
5571 : let logical_sizes_at_once = self
5572 : .conf
5573 : .concurrent_tenant_size_logical_size_queries
5574 : .inner();
5575 :
5576 : // TODO: Having a single mutex block concurrent reads is not great for performance.
5577 : //
5578 : // But the only case where we need to run multiple of these at once is when we
5579 : // request a size for a tenant manually via API, while another background calculation
5580 : // is in progress (which is not a common case).
5581 : //
5582 : // See more for on the issue #2748 condenced out of the initial PR review.
5583 : let mut shared_cache = tokio::select! {
5584 : locked = self.cached_logical_sizes.lock() => locked,
5585 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5586 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
5587 : };
5588 :
5589 : size::gather_inputs(
5590 : self,
5591 : logical_sizes_at_once,
5592 : max_retention_period,
5593 : &mut shared_cache,
5594 : cause,
5595 : cancel,
5596 : ctx,
5597 : )
5598 : .await
5599 : }
5600 :
5601 : /// Calculate synthetic tenant size and cache the result.
5602 : /// This is periodically called by background worker.
5603 : /// result is cached in tenant struct
5604 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
5605 : pub async fn calculate_synthetic_size(
5606 : &self,
5607 : cause: LogicalSizeCalculationCause,
5608 : cancel: &CancellationToken,
5609 : ctx: &RequestContext,
5610 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
5611 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
5612 :
5613 : let size = inputs.calculate();
5614 :
5615 : self.set_cached_synthetic_size(size);
5616 :
5617 : Ok(size)
5618 : }
5619 :
5620 : /// Cache given synthetic size and update the metric value
5621 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
5622 0 : self.cached_synthetic_tenant_size
5623 0 : .store(size, Ordering::Relaxed);
5624 :
5625 : // Only shard zero should be calculating synthetic sizes
5626 0 : debug_assert!(self.shard_identity.is_shard_zero());
5627 :
5628 0 : TENANT_SYNTHETIC_SIZE_METRIC
5629 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
5630 0 : .unwrap()
5631 0 : .set(size);
5632 0 : }
5633 :
5634 0 : pub fn cached_synthetic_size(&self) -> u64 {
5635 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
5636 0 : }
5637 :
5638 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
5639 : ///
5640 : /// This function can take a long time: callers should wrap it in a timeout if calling
5641 : /// from an external API handler.
5642 : ///
5643 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
5644 : /// still bounded by tenant/timeline shutdown.
5645 : #[tracing::instrument(skip_all)]
5646 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
5647 : let timelines = self.timelines.lock().unwrap().clone();
5648 :
5649 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
5650 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
5651 0 : timeline.freeze_and_flush().await?;
5652 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
5653 0 : timeline.remote_client.wait_completion().await?;
5654 :
5655 0 : Ok(())
5656 0 : }
5657 :
5658 : // We do not use a JoinSet for these tasks, because we don't want them to be
5659 : // aborted when this function's future is cancelled: they should stay alive
5660 : // holding their GateGuard until they complete, to ensure their I/Os complete
5661 : // before Timeline shutdown completes.
5662 : let mut results = FuturesUnordered::new();
5663 :
5664 : for (_timeline_id, timeline) in timelines {
5665 : // Run each timeline's flush in a task holding the timeline's gate: this
5666 : // means that if this function's future is cancelled, the Timeline shutdown
5667 : // will still wait for any I/O in here to complete.
5668 : let Ok(gate) = timeline.gate.enter() else {
5669 : continue;
5670 : };
5671 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
5672 : results.push(jh);
5673 : }
5674 :
5675 : while let Some(r) = results.next().await {
5676 : if let Err(e) = r {
5677 : if !e.is_cancelled() && !e.is_panic() {
5678 : tracing::error!("unexpected join error: {e:?}");
5679 : }
5680 : }
5681 : }
5682 :
5683 : // The flushes we did above were just writes, but the TenantShard might have had
5684 : // pending deletions as well from recent compaction/gc: we want to flush those
5685 : // as well. This requires flushing the global delete queue. This is cheap
5686 : // because it's typically a no-op.
5687 : match self.deletion_queue_client.flush_execute().await {
5688 : Ok(_) => {}
5689 : Err(DeletionQueueError::ShuttingDown) => {}
5690 : }
5691 :
5692 : Ok(())
5693 : }
5694 :
5695 0 : pub(crate) fn get_tenant_conf(&self) -> pageserver_api::models::TenantConfig {
5696 0 : self.tenant_conf.load().tenant_conf.clone()
5697 0 : }
5698 :
5699 : /// How much local storage would this tenant like to have? It can cope with
5700 : /// less than this (via eviction and on-demand downloads), but this function enables
5701 : /// the TenantShard to advertise how much storage it would prefer to have to provide fast I/O
5702 : /// by keeping important things on local disk.
5703 : ///
5704 : /// This is a heuristic, not a guarantee: tenants that are long-idle will actually use less
5705 : /// than they report here, due to layer eviction. Tenants with many active branches may
5706 : /// actually use more than they report here.
5707 0 : pub(crate) fn local_storage_wanted(&self) -> u64 {
5708 0 : let timelines = self.timelines.lock().unwrap();
5709 :
5710 : // Heuristic: we use the max() of the timelines' visible sizes, rather than the sum. This
5711 : // reflects the observation that on tenants with multiple large branches, typically only one
5712 : // of them is used actively enough to occupy space on disk.
5713 0 : timelines
5714 0 : .values()
5715 0 : .map(|t| t.metrics.visible_physical_size_gauge.get())
5716 0 : .max()
5717 0 : .unwrap_or(0)
5718 0 : }
5719 :
5720 : /// Builds a new tenant manifest, and uploads it if it differs from the last-known tenant
5721 : /// manifest in `Self::remote_tenant_manifest`.
5722 : ///
5723 : /// TODO: instead of requiring callers to remember to call `maybe_upload_tenant_manifest` after
5724 : /// changing any `TenantShard` state that's included in the manifest, consider making the manifest
5725 : /// the authoritative source of data with an API that automatically uploads on changes. Revisit
5726 : /// this when the manifest is more widely used and we have a better idea of the data model.
5727 119 : pub(crate) async fn maybe_upload_tenant_manifest(&self) -> Result<(), TenantManifestError> {
5728 : // Multiple tasks may call this function concurrently after mutating the TenantShard runtime
5729 : // state, affecting the manifest generated by `build_tenant_manifest`. We use an async mutex
5730 : // to serialize these callers. `eq_ignoring_version` acts as a slightly inefficient but
5731 : // simple coalescing mechanism.
5732 119 : let mut guard = tokio::select! {
5733 119 : guard = self.remote_tenant_manifest.lock() => guard,
5734 119 : _ = self.cancel.cancelled() => return Err(TenantManifestError::Cancelled),
5735 : };
5736 :
5737 : // Build a new manifest.
5738 119 : let manifest = self.build_tenant_manifest();
5739 :
5740 : // Check if the manifest has changed. We ignore the version number here, to avoid
5741 : // uploading every manifest on version number bumps.
5742 119 : if let Some(old) = guard.as_ref() {
5743 4 : if manifest.eq_ignoring_version(old) {
5744 3 : return Ok(());
5745 1 : }
5746 115 : }
5747 :
5748 : // Update metrics
5749 116 : let tid = self.tenant_shard_id.to_string();
5750 116 : let shard_id = self.tenant_shard_id.shard_slug().to_string();
5751 116 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
5752 116 : TENANT_OFFLOADED_TIMELINES
5753 116 : .with_label_values(set_key)
5754 116 : .set(manifest.offloaded_timelines.len() as u64);
5755 :
5756 : // Upload the manifest. Remote storage does no retries internally, so retry here.
5757 116 : match backoff::retry(
5758 116 : || async {
5759 116 : upload_tenant_manifest(
5760 116 : &self.remote_storage,
5761 116 : &self.tenant_shard_id,
5762 116 : self.generation,
5763 116 : &manifest,
5764 116 : &self.cancel,
5765 116 : )
5766 116 : .await
5767 232 : },
5768 0 : |_| self.cancel.is_cancelled(),
5769 : FAILED_UPLOAD_WARN_THRESHOLD,
5770 : FAILED_REMOTE_OP_RETRIES,
5771 116 : "uploading tenant manifest",
5772 116 : &self.cancel,
5773 : )
5774 116 : .await
5775 : {
5776 0 : None => Err(TenantManifestError::Cancelled),
5777 0 : Some(Err(_)) if self.cancel.is_cancelled() => Err(TenantManifestError::Cancelled),
5778 0 : Some(Err(e)) => Err(TenantManifestError::RemoteStorage(e)),
5779 : Some(Ok(_)) => {
5780 : // Store the successfully uploaded manifest, so that future callers can avoid
5781 : // re-uploading the same thing.
5782 116 : *guard = Some(manifest);
5783 :
5784 116 : Ok(())
5785 : }
5786 : }
5787 119 : }
5788 : }
5789 :
5790 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
5791 : /// to get bootstrap data for timeline initialization.
5792 0 : async fn run_initdb(
5793 0 : conf: &'static PageServerConf,
5794 0 : initdb_target_dir: &Utf8Path,
5795 0 : pg_version: PgMajorVersion,
5796 0 : cancel: &CancellationToken,
5797 0 : ) -> Result<(), InitdbError> {
5798 0 : let initdb_bin_path = conf
5799 0 : .pg_bin_dir(pg_version)
5800 0 : .map_err(InitdbError::Other)?
5801 0 : .join("initdb");
5802 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
5803 0 : info!(
5804 0 : "running {} in {}, libdir: {}",
5805 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
5806 : );
5807 :
5808 0 : let _permit = {
5809 0 : let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
5810 0 : INIT_DB_SEMAPHORE.acquire().await
5811 : };
5812 :
5813 0 : CONCURRENT_INITDBS.inc();
5814 0 : scopeguard::defer! {
5815 : CONCURRENT_INITDBS.dec();
5816 : }
5817 :
5818 0 : let _timer = INITDB_RUN_TIME.start_timer();
5819 0 : let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
5820 0 : superuser: &conf.superuser,
5821 0 : locale: &conf.locale,
5822 0 : initdb_bin: &initdb_bin_path,
5823 0 : pg_version,
5824 0 : library_search_path: &initdb_lib_dir,
5825 0 : pgdata: initdb_target_dir,
5826 0 : })
5827 0 : .await
5828 0 : .map_err(InitdbError::Inner);
5829 :
5830 : // This isn't true cancellation support, see above. Still return an error to
5831 : // excercise the cancellation code path.
5832 0 : if cancel.is_cancelled() {
5833 0 : return Err(InitdbError::Cancelled);
5834 0 : }
5835 :
5836 0 : res
5837 0 : }
5838 :
5839 : /// Dump contents of a layer file to stdout.
5840 0 : pub async fn dump_layerfile_from_path(
5841 0 : path: &Utf8Path,
5842 0 : verbose: bool,
5843 0 : ctx: &RequestContext,
5844 0 : ) -> anyhow::Result<()> {
5845 : use std::os::unix::fs::FileExt;
5846 :
5847 : // All layer files start with a two-byte "magic" value, to identify the kind of
5848 : // file.
5849 0 : let file = File::open(path)?;
5850 0 : let mut header_buf = [0u8; 2];
5851 0 : file.read_exact_at(&mut header_buf, 0)?;
5852 :
5853 0 : match u16::from_be_bytes(header_buf) {
5854 : crate::IMAGE_FILE_MAGIC => {
5855 0 : ImageLayer::new_for_path(path, file)?
5856 0 : .dump(verbose, ctx)
5857 0 : .await?
5858 : }
5859 : crate::DELTA_FILE_MAGIC => {
5860 0 : DeltaLayer::new_for_path(path, file)?
5861 0 : .dump(verbose, ctx)
5862 0 : .await?
5863 : }
5864 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
5865 : }
5866 :
5867 0 : Ok(())
5868 0 : }
5869 :
5870 : #[cfg(test)]
5871 : pub(crate) mod harness {
5872 : use bytes::{Bytes, BytesMut};
5873 : use hex_literal::hex;
5874 : use once_cell::sync::OnceCell;
5875 : use pageserver_api::key::Key;
5876 : use pageserver_api::models::ShardParameters;
5877 : use pageserver_api::shard::ShardIndex;
5878 : use utils::id::TenantId;
5879 : use utils::logging;
5880 : use wal_decoder::models::record::NeonWalRecord;
5881 :
5882 : use super::*;
5883 : use crate::deletion_queue::mock::MockDeletionQueue;
5884 : use crate::l0_flush::L0FlushConfig;
5885 : use crate::walredo::apply_neon;
5886 :
5887 : pub const TIMELINE_ID: TimelineId =
5888 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
5889 : pub const NEW_TIMELINE_ID: TimelineId =
5890 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
5891 :
5892 : /// Convenience function to create a page image with given string as the only content
5893 2514377 : pub fn test_img(s: &str) -> Bytes {
5894 2514377 : let mut buf = BytesMut::new();
5895 2514377 : buf.extend_from_slice(s.as_bytes());
5896 2514377 : buf.resize(64, 0);
5897 :
5898 2514377 : buf.freeze()
5899 2514377 : }
5900 :
5901 : pub struct TenantHarness {
5902 : pub conf: &'static PageServerConf,
5903 : pub tenant_conf: pageserver_api::models::TenantConfig,
5904 : pub tenant_shard_id: TenantShardId,
5905 : pub shard_identity: ShardIdentity,
5906 : pub generation: Generation,
5907 : pub shard: ShardIndex,
5908 : pub remote_storage: GenericRemoteStorage,
5909 : pub remote_fs_dir: Utf8PathBuf,
5910 : pub deletion_queue: MockDeletionQueue,
5911 : }
5912 :
5913 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
5914 :
5915 130 : pub(crate) fn setup_logging() {
5916 130 : LOG_HANDLE.get_or_init(|| {
5917 124 : logging::init(
5918 124 : logging::LogFormat::Test,
5919 : // enable it in case the tests exercise code paths that use
5920 : // debug_assert_current_span_has_tenant_and_timeline_id
5921 124 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
5922 124 : logging::Output::Stdout,
5923 : )
5924 124 : .expect("Failed to init test logging");
5925 124 : });
5926 130 : }
5927 :
5928 : impl TenantHarness {
5929 118 : pub async fn create_custom(
5930 118 : test_name: &'static str,
5931 118 : tenant_conf: pageserver_api::models::TenantConfig,
5932 118 : tenant_id: TenantId,
5933 118 : shard_identity: ShardIdentity,
5934 118 : generation: Generation,
5935 118 : ) -> anyhow::Result<Self> {
5936 118 : setup_logging();
5937 :
5938 118 : let repo_dir = PageServerConf::test_repo_dir(test_name);
5939 118 : let _ = fs::remove_dir_all(&repo_dir);
5940 118 : fs::create_dir_all(&repo_dir)?;
5941 :
5942 118 : let conf = PageServerConf::dummy_conf(repo_dir);
5943 : // Make a static copy of the config. This can never be free'd, but that's
5944 : // OK in a test.
5945 118 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
5946 :
5947 118 : let shard = shard_identity.shard_index();
5948 118 : let tenant_shard_id = TenantShardId {
5949 118 : tenant_id,
5950 118 : shard_number: shard.shard_number,
5951 118 : shard_count: shard.shard_count,
5952 118 : };
5953 118 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
5954 118 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
5955 :
5956 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
5957 118 : let remote_fs_dir = conf.workdir.join("localfs");
5958 118 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
5959 118 : let config = RemoteStorageConfig {
5960 118 : storage: RemoteStorageKind::LocalFs {
5961 118 : local_path: remote_fs_dir.clone(),
5962 118 : },
5963 118 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
5964 118 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
5965 118 : };
5966 118 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
5967 118 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
5968 :
5969 118 : Ok(Self {
5970 118 : conf,
5971 118 : tenant_conf,
5972 118 : tenant_shard_id,
5973 118 : shard_identity,
5974 118 : generation,
5975 118 : shard,
5976 118 : remote_storage,
5977 118 : remote_fs_dir,
5978 118 : deletion_queue,
5979 118 : })
5980 118 : }
5981 :
5982 110 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
5983 : // Disable automatic GC and compaction to make the unit tests more deterministic.
5984 : // The tests perform them manually if needed.
5985 110 : let tenant_conf = pageserver_api::models::TenantConfig {
5986 110 : gc_period: Some(Duration::ZERO),
5987 110 : compaction_period: Some(Duration::ZERO),
5988 110 : ..Default::default()
5989 110 : };
5990 110 : let tenant_id = TenantId::generate();
5991 110 : let shard = ShardIdentity::unsharded();
5992 110 : Self::create_custom(
5993 110 : test_name,
5994 110 : tenant_conf,
5995 110 : tenant_id,
5996 110 : shard,
5997 110 : Generation::new(0xdeadbeef),
5998 110 : )
5999 110 : .await
6000 110 : }
6001 :
6002 10 : pub fn span(&self) -> tracing::Span {
6003 10 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
6004 10 : }
6005 :
6006 118 : pub(crate) async fn load(&self) -> (Arc<TenantShard>, RequestContext) {
6007 118 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error)
6008 118 : .with_scope_unit_test();
6009 : (
6010 118 : self.do_try_load(&ctx)
6011 118 : .await
6012 118 : .expect("failed to load test tenant"),
6013 118 : ctx,
6014 : )
6015 118 : }
6016 :
6017 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
6018 : pub(crate) async fn do_try_load(
6019 : &self,
6020 : ctx: &RequestContext,
6021 : ) -> anyhow::Result<Arc<TenantShard>> {
6022 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
6023 :
6024 : let (basebackup_cache, _) = BasebackupCache::new(Utf8PathBuf::new(), None);
6025 :
6026 : let tenant = Arc::new(TenantShard::new(
6027 : TenantState::Attaching,
6028 : self.conf,
6029 : AttachedTenantConf::try_from(
6030 : self.conf,
6031 : LocationConf::attached_single(
6032 : self.tenant_conf.clone(),
6033 : self.generation,
6034 : ShardParameters::default(),
6035 : ),
6036 : )
6037 : .unwrap(),
6038 : self.shard_identity,
6039 : Some(walredo_mgr),
6040 : self.tenant_shard_id,
6041 : self.remote_storage.clone(),
6042 : self.deletion_queue.new_client(),
6043 : // TODO: ideally we should run all unit tests with both configs
6044 : L0FlushGlobalState::new(L0FlushConfig::default()),
6045 : basebackup_cache,
6046 : FeatureResolver::new_disabled(),
6047 : ));
6048 :
6049 : let preload = tenant
6050 : .preload(&self.remote_storage, CancellationToken::new())
6051 : .await?;
6052 : tenant.attach(Some(preload), ctx).await?;
6053 :
6054 : tenant.state.send_replace(TenantState::Active);
6055 : for timeline in tenant.timelines.lock().unwrap().values() {
6056 : timeline.set_state(TimelineState::Active);
6057 : }
6058 : Ok(tenant)
6059 : }
6060 :
6061 1 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
6062 1 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
6063 1 : }
6064 : }
6065 :
6066 : // Mock WAL redo manager that doesn't do much
6067 : pub(crate) struct TestRedoManager;
6068 :
6069 : impl TestRedoManager {
6070 : /// # Cancel-Safety
6071 : ///
6072 : /// This method is cancellation-safe.
6073 26774 : pub async fn request_redo(
6074 26774 : &self,
6075 26774 : key: Key,
6076 26774 : lsn: Lsn,
6077 26774 : base_img: Option<(Lsn, Bytes)>,
6078 26774 : records: Vec<(Lsn, NeonWalRecord)>,
6079 26774 : _pg_version: PgMajorVersion,
6080 26774 : _redo_attempt_type: RedoAttemptType,
6081 26774 : ) -> Result<Bytes, walredo::Error> {
6082 1403510 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
6083 26774 : if records_neon {
6084 : // For Neon wal records, we can decode without spawning postgres, so do so.
6085 26774 : let mut page = match (base_img, records.first()) {
6086 13029 : (Some((_lsn, img)), _) => {
6087 13029 : let mut page = BytesMut::new();
6088 13029 : page.extend_from_slice(&img);
6089 13029 : page
6090 : }
6091 13745 : (_, Some((_lsn, rec))) if rec.will_init() => BytesMut::new(),
6092 : _ => {
6093 0 : panic!("Neon WAL redo requires base image or will init record");
6094 : }
6095 : };
6096 :
6097 1430283 : for (record_lsn, record) in records {
6098 1403510 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
6099 : }
6100 26773 : Ok(page.freeze())
6101 : } else {
6102 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
6103 0 : let s = format!(
6104 0 : "redo for {} to get to {}, with {} and {} records",
6105 : key,
6106 : lsn,
6107 0 : if base_img.is_some() {
6108 0 : "base image"
6109 : } else {
6110 0 : "no base image"
6111 : },
6112 0 : records.len()
6113 : );
6114 0 : println!("{s}");
6115 :
6116 0 : Ok(test_img(&s))
6117 : }
6118 26774 : }
6119 : }
6120 : }
6121 :
6122 : #[cfg(test)]
6123 : mod tests {
6124 : use std::collections::{BTreeMap, BTreeSet};
6125 :
6126 : use bytes::{Bytes, BytesMut};
6127 : use hex_literal::hex;
6128 : use itertools::Itertools;
6129 : #[cfg(feature = "testing")]
6130 : use models::CompactLsnRange;
6131 : use pageserver_api::key::{
6132 : AUX_KEY_PREFIX, Key, NON_INHERITED_RANGE, RELATION_SIZE_PREFIX, repl_origin_key,
6133 : };
6134 : use pageserver_api::keyspace::KeySpace;
6135 : #[cfg(feature = "testing")]
6136 : use pageserver_api::keyspace::KeySpaceRandomAccum;
6137 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings, LsnLease};
6138 : use pageserver_compaction::helpers::overlaps_with;
6139 : #[cfg(feature = "testing")]
6140 : use rand::SeedableRng;
6141 : #[cfg(feature = "testing")]
6142 : use rand::rngs::StdRng;
6143 : use rand::{Rng, thread_rng};
6144 : #[cfg(feature = "testing")]
6145 : use std::ops::Range;
6146 : use storage_layer::{IoConcurrency, PersistentLayerKey};
6147 : use tests::storage_layer::ValuesReconstructState;
6148 : use tests::timeline::{GetVectoredError, ShutdownMode};
6149 : #[cfg(feature = "testing")]
6150 : use timeline::GcInfo;
6151 : #[cfg(feature = "testing")]
6152 : use timeline::InMemoryLayerTestDesc;
6153 : #[cfg(feature = "testing")]
6154 : use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
6155 : use timeline::{CompactOptions, DeltaLayerTestDesc, VersionedKeySpaceQuery};
6156 : use utils::id::TenantId;
6157 : use utils::shard::{ShardCount, ShardNumber};
6158 : #[cfg(feature = "testing")]
6159 : use wal_decoder::models::record::NeonWalRecord;
6160 : use wal_decoder::models::value::Value;
6161 :
6162 : use super::*;
6163 : use crate::DEFAULT_PG_VERSION;
6164 : use crate::keyspace::KeySpaceAccum;
6165 : use crate::tenant::harness::*;
6166 : use crate::tenant::timeline::CompactFlags;
6167 :
6168 : static TEST_KEY: Lazy<Key> =
6169 10 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
6170 :
6171 : #[cfg(feature = "testing")]
6172 : struct TestTimelineSpecification {
6173 : start_lsn: Lsn,
6174 : last_record_lsn: Lsn,
6175 :
6176 : in_memory_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
6177 : delta_layers_shape: Vec<(Range<Key>, Range<Lsn>)>,
6178 : image_layers_shape: Vec<(Range<Key>, Lsn)>,
6179 :
6180 : gap_chance: u8,
6181 : will_init_chance: u8,
6182 : }
6183 :
6184 : #[cfg(feature = "testing")]
6185 : struct Storage {
6186 : storage: HashMap<(Key, Lsn), Value>,
6187 : start_lsn: Lsn,
6188 : }
6189 :
6190 : #[cfg(feature = "testing")]
6191 : impl Storage {
6192 32000 : fn get(&self, key: Key, lsn: Lsn) -> Bytes {
6193 : use bytes::BufMut;
6194 :
6195 32000 : let mut crnt_lsn = lsn;
6196 32000 : let mut got_base = false;
6197 :
6198 32000 : let mut acc = Vec::new();
6199 :
6200 2831871 : while crnt_lsn >= self.start_lsn {
6201 2831871 : if let Some(value) = self.storage.get(&(key, crnt_lsn)) {
6202 1421172 : acc.push(value.clone());
6203 :
6204 1402881 : match value {
6205 1402881 : Value::WalRecord(NeonWalRecord::Test { will_init, .. }) => {
6206 1402881 : if *will_init {
6207 13709 : got_base = true;
6208 13709 : break;
6209 1389172 : }
6210 : }
6211 : Value::Image(_) => {
6212 18291 : got_base = true;
6213 18291 : break;
6214 : }
6215 0 : _ => unreachable!(),
6216 : }
6217 1410699 : }
6218 :
6219 2799871 : crnt_lsn = crnt_lsn.checked_sub(1u64).unwrap();
6220 : }
6221 :
6222 32000 : assert!(
6223 32000 : got_base,
6224 0 : "Input data was incorrect. No base image for {key}@{lsn}"
6225 : );
6226 :
6227 32000 : tracing::debug!("Wal redo depth for {key}@{lsn} is {}", acc.len());
6228 :
6229 32000 : let mut blob = BytesMut::new();
6230 1421172 : for value in acc.into_iter().rev() {
6231 1402881 : match value {
6232 1402881 : Value::WalRecord(NeonWalRecord::Test { append, .. }) => {
6233 1402881 : blob.extend_from_slice(append.as_bytes());
6234 1402881 : }
6235 18291 : Value::Image(img) => {
6236 18291 : blob.put(img);
6237 18291 : }
6238 0 : _ => unreachable!(),
6239 : }
6240 : }
6241 :
6242 32000 : blob.into()
6243 32000 : }
6244 : }
6245 :
6246 : #[cfg(feature = "testing")]
6247 : #[allow(clippy::too_many_arguments)]
6248 1 : async fn randomize_timeline(
6249 1 : tenant: &Arc<TenantShard>,
6250 1 : new_timeline_id: TimelineId,
6251 1 : pg_version: PgMajorVersion,
6252 1 : spec: TestTimelineSpecification,
6253 1 : random: &mut rand::rngs::StdRng,
6254 1 : ctx: &RequestContext,
6255 1 : ) -> anyhow::Result<(Arc<Timeline>, Storage, Vec<Lsn>)> {
6256 1 : let mut storage: HashMap<(Key, Lsn), Value> = HashMap::default();
6257 1 : let mut interesting_lsns = vec![spec.last_record_lsn];
6258 :
6259 2 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6260 2 : let mut lsn = lsn_range.start;
6261 202 : while lsn < lsn_range.end {
6262 200 : let mut key = key_range.start;
6263 21018 : while key < key_range.end {
6264 20818 : let gap = random.gen_range(1..=100) <= spec.gap_chance;
6265 20818 : let will_init = random.gen_range(1..=100) <= spec.will_init_chance;
6266 :
6267 20818 : if gap {
6268 1018 : continue;
6269 19800 : }
6270 :
6271 19800 : let record = if will_init {
6272 191 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6273 : } else {
6274 19609 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6275 : };
6276 :
6277 19800 : storage.insert((key, lsn), record);
6278 :
6279 19800 : key = key.next();
6280 : }
6281 200 : lsn = Lsn(lsn.0 + 1);
6282 : }
6283 :
6284 : // Stash some interesting LSN for future use
6285 6 : for offset in [0, 5, 100].iter() {
6286 6 : if *offset == 0 {
6287 2 : interesting_lsns.push(lsn_range.start);
6288 2 : } else {
6289 4 : let below = lsn_range.start.checked_sub(*offset);
6290 4 : match below {
6291 4 : Some(v) if v >= spec.start_lsn => {
6292 4 : interesting_lsns.push(v);
6293 4 : }
6294 0 : _ => {}
6295 : }
6296 :
6297 4 : let above = Lsn(lsn_range.start.0 + offset);
6298 4 : interesting_lsns.push(above);
6299 : }
6300 : }
6301 : }
6302 :
6303 3 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6304 3 : let mut lsn = lsn_range.start;
6305 315 : while lsn < lsn_range.end {
6306 312 : let mut key = key_range.start;
6307 11112 : while key < key_range.end {
6308 10800 : let gap = random.gen_range(1..=100) <= spec.gap_chance;
6309 10800 : let will_init = random.gen_range(1..=100) <= spec.will_init_chance;
6310 :
6311 10800 : if gap {
6312 504 : continue;
6313 10296 : }
6314 :
6315 10296 : let record = if will_init {
6316 103 : Value::WalRecord(NeonWalRecord::wal_init(format!("[wil_init {key}@{lsn}]")))
6317 : } else {
6318 10193 : Value::WalRecord(NeonWalRecord::wal_append(format!("[delta {key}@{lsn}]")))
6319 : };
6320 :
6321 10296 : storage.insert((key, lsn), record);
6322 :
6323 10296 : key = key.next();
6324 : }
6325 312 : lsn = Lsn(lsn.0 + 1);
6326 : }
6327 :
6328 : // Stash some interesting LSN for future use
6329 9 : for offset in [0, 5, 100].iter() {
6330 9 : if *offset == 0 {
6331 3 : interesting_lsns.push(lsn_range.start);
6332 3 : } else {
6333 6 : let below = lsn_range.start.checked_sub(*offset);
6334 6 : match below {
6335 6 : Some(v) if v >= spec.start_lsn => {
6336 3 : interesting_lsns.push(v);
6337 3 : }
6338 3 : _ => {}
6339 : }
6340 :
6341 6 : let above = Lsn(lsn_range.start.0 + offset);
6342 6 : interesting_lsns.push(above);
6343 : }
6344 : }
6345 : }
6346 :
6347 3 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6348 3 : let mut key = key_range.start;
6349 142 : while key < key_range.end {
6350 139 : let blob = Bytes::from(format!("[image {key}@{lsn}]"));
6351 139 : let record = Value::Image(blob.clone());
6352 139 : storage.insert((key, *lsn), record);
6353 139 :
6354 139 : key = key.next();
6355 139 : }
6356 :
6357 : // Stash some interesting LSN for future use
6358 9 : for offset in [0, 5, 100].iter() {
6359 9 : if *offset == 0 {
6360 3 : interesting_lsns.push(*lsn);
6361 3 : } else {
6362 6 : let below = lsn.checked_sub(*offset);
6363 6 : match below {
6364 6 : Some(v) if v >= spec.start_lsn => {
6365 4 : interesting_lsns.push(v);
6366 4 : }
6367 2 : _ => {}
6368 : }
6369 :
6370 6 : let above = Lsn(lsn.0 + offset);
6371 6 : interesting_lsns.push(above);
6372 : }
6373 : }
6374 : }
6375 :
6376 1 : let in_memory_test_layers = {
6377 1 : let mut acc = Vec::new();
6378 :
6379 2 : for (key_range, lsn_range) in spec.in_memory_layers_shape.iter() {
6380 2 : let mut data = Vec::new();
6381 :
6382 2 : let mut lsn = lsn_range.start;
6383 202 : while lsn < lsn_range.end {
6384 200 : let mut key = key_range.start;
6385 20000 : while key < key_range.end {
6386 19800 : if let Some(record) = storage.get(&(key, lsn)) {
6387 19800 : data.push((key, lsn, record.clone()));
6388 19800 : }
6389 :
6390 19800 : key = key.next();
6391 : }
6392 200 : lsn = Lsn(lsn.0 + 1);
6393 : }
6394 :
6395 2 : acc.push(InMemoryLayerTestDesc {
6396 2 : data,
6397 2 : lsn_range: lsn_range.clone(),
6398 2 : is_open: false,
6399 2 : })
6400 : }
6401 :
6402 1 : acc
6403 : };
6404 :
6405 1 : let delta_test_layers = {
6406 1 : let mut acc = Vec::new();
6407 :
6408 3 : for (key_range, lsn_range) in spec.delta_layers_shape.iter() {
6409 3 : let mut data = Vec::new();
6410 :
6411 3 : let mut lsn = lsn_range.start;
6412 315 : while lsn < lsn_range.end {
6413 312 : let mut key = key_range.start;
6414 10608 : while key < key_range.end {
6415 10296 : if let Some(record) = storage.get(&(key, lsn)) {
6416 10296 : data.push((key, lsn, record.clone()));
6417 10296 : }
6418 :
6419 10296 : key = key.next();
6420 : }
6421 312 : lsn = Lsn(lsn.0 + 1);
6422 : }
6423 :
6424 3 : acc.push(DeltaLayerTestDesc {
6425 3 : data,
6426 3 : lsn_range: lsn_range.clone(),
6427 3 : key_range: key_range.clone(),
6428 3 : })
6429 : }
6430 :
6431 1 : acc
6432 : };
6433 :
6434 1 : let image_test_layers = {
6435 1 : let mut acc = Vec::new();
6436 :
6437 3 : for (key_range, lsn) in spec.image_layers_shape.iter() {
6438 3 : let mut data = Vec::new();
6439 :
6440 3 : let mut key = key_range.start;
6441 142 : while key < key_range.end {
6442 139 : if let Some(record) = storage.get(&(key, *lsn)) {
6443 139 : let blob = match record {
6444 139 : Value::Image(blob) => blob.clone(),
6445 0 : _ => unreachable!(),
6446 : };
6447 :
6448 139 : data.push((key, blob));
6449 0 : }
6450 :
6451 139 : key = key.next();
6452 : }
6453 :
6454 3 : acc.push((*lsn, data));
6455 : }
6456 :
6457 1 : acc
6458 : };
6459 :
6460 1 : let tline = tenant
6461 1 : .create_test_timeline_with_layers(
6462 1 : new_timeline_id,
6463 1 : spec.start_lsn,
6464 1 : pg_version,
6465 1 : ctx,
6466 1 : in_memory_test_layers,
6467 1 : delta_test_layers,
6468 1 : image_test_layers,
6469 1 : spec.last_record_lsn,
6470 1 : )
6471 1 : .await?;
6472 :
6473 1 : Ok((
6474 1 : tline,
6475 1 : Storage {
6476 1 : storage,
6477 1 : start_lsn: spec.start_lsn,
6478 1 : },
6479 1 : interesting_lsns,
6480 1 : ))
6481 1 : }
6482 :
6483 : #[tokio::test]
6484 1 : async fn test_basic() -> anyhow::Result<()> {
6485 1 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
6486 1 : let tline = tenant
6487 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
6488 1 : .await?;
6489 :
6490 1 : let mut writer = tline.writer().await;
6491 1 : writer
6492 1 : .put(
6493 1 : *TEST_KEY,
6494 1 : Lsn(0x10),
6495 1 : &Value::Image(test_img("foo at 0x10")),
6496 1 : &ctx,
6497 1 : )
6498 1 : .await?;
6499 1 : writer.finish_write(Lsn(0x10));
6500 1 : drop(writer);
6501 :
6502 1 : let mut writer = tline.writer().await;
6503 1 : writer
6504 1 : .put(
6505 1 : *TEST_KEY,
6506 1 : Lsn(0x20),
6507 1 : &Value::Image(test_img("foo at 0x20")),
6508 1 : &ctx,
6509 1 : )
6510 1 : .await?;
6511 1 : writer.finish_write(Lsn(0x20));
6512 1 : drop(writer);
6513 :
6514 1 : assert_eq!(
6515 1 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
6516 1 : test_img("foo at 0x10")
6517 : );
6518 1 : assert_eq!(
6519 1 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
6520 1 : test_img("foo at 0x10")
6521 : );
6522 1 : assert_eq!(
6523 1 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
6524 1 : test_img("foo at 0x20")
6525 : );
6526 :
6527 2 : Ok(())
6528 1 : }
6529 :
6530 : #[tokio::test]
6531 1 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
6532 1 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
6533 1 : .await?
6534 1 : .load()
6535 1 : .await;
6536 1 : let _ = tenant
6537 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6538 1 : .await?;
6539 :
6540 1 : match tenant
6541 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6542 1 : .await
6543 1 : {
6544 1 : Ok(_) => panic!("duplicate timeline creation should fail"),
6545 1 : Err(e) => assert_eq!(
6546 1 : e.to_string(),
6547 1 : "timeline already exists with different parameters".to_string()
6548 1 : ),
6549 1 : }
6550 1 :
6551 1 : Ok(())
6552 1 : }
6553 :
6554 : /// Convenience function to create a page image with given string as the only content
6555 5 : pub fn test_value(s: &str) -> Value {
6556 5 : let mut buf = BytesMut::new();
6557 5 : buf.extend_from_slice(s.as_bytes());
6558 5 : Value::Image(buf.freeze())
6559 5 : }
6560 :
6561 : ///
6562 : /// Test branch creation
6563 : ///
6564 : #[tokio::test]
6565 1 : async fn test_branch() -> anyhow::Result<()> {
6566 : use std::str::from_utf8;
6567 :
6568 1 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
6569 1 : let tline = tenant
6570 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6571 1 : .await?;
6572 1 : let mut writer = tline.writer().await;
6573 :
6574 : #[allow(non_snake_case)]
6575 1 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
6576 : #[allow(non_snake_case)]
6577 1 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
6578 :
6579 : // Insert a value on the timeline
6580 1 : writer
6581 1 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
6582 1 : .await?;
6583 1 : writer
6584 1 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
6585 1 : .await?;
6586 1 : writer.finish_write(Lsn(0x20));
6587 :
6588 1 : writer
6589 1 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
6590 1 : .await?;
6591 1 : writer.finish_write(Lsn(0x30));
6592 1 : writer
6593 1 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
6594 1 : .await?;
6595 1 : writer.finish_write(Lsn(0x40));
6596 :
6597 : //assert_current_logical_size(&tline, Lsn(0x40));
6598 :
6599 : // Branch the history, modify relation differently on the new timeline
6600 1 : tenant
6601 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
6602 1 : .await?;
6603 1 : let newtline = tenant
6604 1 : .get_timeline(NEW_TIMELINE_ID, true)
6605 1 : .expect("Should have a local timeline");
6606 1 : let mut new_writer = newtline.writer().await;
6607 1 : new_writer
6608 1 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
6609 1 : .await?;
6610 1 : new_writer.finish_write(Lsn(0x40));
6611 :
6612 : // Check page contents on both branches
6613 1 : assert_eq!(
6614 1 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6615 : "foo at 0x40"
6616 : );
6617 1 : assert_eq!(
6618 1 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
6619 : "bar at 0x40"
6620 : );
6621 1 : assert_eq!(
6622 1 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
6623 : "foobar at 0x20"
6624 : );
6625 :
6626 : //assert_current_logical_size(&tline, Lsn(0x40));
6627 :
6628 2 : Ok(())
6629 1 : }
6630 :
6631 10 : async fn make_some_layers(
6632 10 : tline: &Timeline,
6633 10 : start_lsn: Lsn,
6634 10 : ctx: &RequestContext,
6635 10 : ) -> anyhow::Result<()> {
6636 10 : let mut lsn = start_lsn;
6637 : {
6638 10 : let mut writer = tline.writer().await;
6639 : // Create a relation on the timeline
6640 10 : writer
6641 10 : .put(
6642 10 : *TEST_KEY,
6643 10 : lsn,
6644 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6645 10 : ctx,
6646 10 : )
6647 10 : .await?;
6648 10 : writer.finish_write(lsn);
6649 10 : lsn += 0x10;
6650 10 : writer
6651 10 : .put(
6652 10 : *TEST_KEY,
6653 10 : lsn,
6654 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6655 10 : ctx,
6656 10 : )
6657 10 : .await?;
6658 10 : writer.finish_write(lsn);
6659 10 : lsn += 0x10;
6660 : }
6661 10 : tline.freeze_and_flush().await?;
6662 : {
6663 10 : let mut writer = tline.writer().await;
6664 10 : writer
6665 10 : .put(
6666 10 : *TEST_KEY,
6667 10 : lsn,
6668 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6669 10 : ctx,
6670 10 : )
6671 10 : .await?;
6672 10 : writer.finish_write(lsn);
6673 10 : lsn += 0x10;
6674 10 : writer
6675 10 : .put(
6676 10 : *TEST_KEY,
6677 10 : lsn,
6678 10 : &Value::Image(test_img(&format!("foo at {lsn}"))),
6679 10 : ctx,
6680 10 : )
6681 10 : .await?;
6682 10 : writer.finish_write(lsn);
6683 : }
6684 10 : tline.freeze_and_flush().await.map_err(|e| e.into())
6685 10 : }
6686 :
6687 : #[tokio::test]
6688 1 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
6689 1 : let (tenant, ctx) =
6690 1 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
6691 1 : .await?
6692 1 : .load()
6693 1 : .await;
6694 1 : let tline = tenant
6695 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6696 1 : .await?;
6697 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6698 :
6699 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6700 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
6701 : // and compaction works. But it does set the 'cutoff' point so that the cross check
6702 : // below should fail.
6703 1 : tenant
6704 1 : .gc_iteration(
6705 1 : Some(TIMELINE_ID),
6706 1 : 0x10,
6707 1 : Duration::ZERO,
6708 1 : &CancellationToken::new(),
6709 1 : &ctx,
6710 1 : )
6711 1 : .await?;
6712 :
6713 : // try to branch at lsn 25, should fail because we already garbage collected the data
6714 1 : match tenant
6715 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6716 1 : .await
6717 1 : {
6718 1 : Ok(_) => panic!("branching should have failed"),
6719 1 : Err(err) => {
6720 1 : let CreateTimelineError::AncestorLsn(err) = err else {
6721 1 : panic!("wrong error type")
6722 1 : };
6723 1 : assert!(err.to_string().contains("invalid branch start lsn"));
6724 1 : assert!(
6725 1 : err.source()
6726 1 : .unwrap()
6727 1 : .to_string()
6728 1 : .contains("we might've already garbage collected needed data")
6729 1 : )
6730 1 : }
6731 1 : }
6732 1 :
6733 1 : Ok(())
6734 1 : }
6735 :
6736 : #[tokio::test]
6737 1 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
6738 1 : let (tenant, ctx) =
6739 1 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
6740 1 : .await?
6741 1 : .load()
6742 1 : .await;
6743 :
6744 1 : let tline = tenant
6745 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
6746 1 : .await?;
6747 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
6748 1 : match tenant
6749 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
6750 1 : .await
6751 1 : {
6752 1 : Ok(_) => panic!("branching should have failed"),
6753 1 : Err(err) => {
6754 1 : let CreateTimelineError::AncestorLsn(err) = err else {
6755 1 : panic!("wrong error type");
6756 1 : };
6757 1 : assert!(&err.to_string().contains("invalid branch start lsn"));
6758 1 : assert!(
6759 1 : &err.source()
6760 1 : .unwrap()
6761 1 : .to_string()
6762 1 : .contains("is earlier than latest GC cutoff")
6763 1 : );
6764 1 : }
6765 1 : }
6766 1 :
6767 1 : Ok(())
6768 1 : }
6769 :
6770 : /*
6771 : // FIXME: This currently fails to error out. Calling GC doesn't currently
6772 : // remove the old value, we'd need to work a little harder
6773 : #[tokio::test]
6774 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
6775 : let repo =
6776 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
6777 : .load();
6778 :
6779 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
6780 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6781 :
6782 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
6783 : let applied_gc_cutoff_lsn = tline.get_applied_gc_cutoff_lsn();
6784 : assert!(*applied_gc_cutoff_lsn > Lsn(0x25));
6785 : match tline.get(*TEST_KEY, Lsn(0x25)) {
6786 : Ok(_) => panic!("request for page should have failed"),
6787 : Err(err) => assert!(err.to_string().contains("not found at")),
6788 : }
6789 : Ok(())
6790 : }
6791 : */
6792 :
6793 : #[tokio::test]
6794 1 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
6795 1 : let (tenant, ctx) =
6796 1 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
6797 1 : .await?
6798 1 : .load()
6799 1 : .await;
6800 1 : let tline = tenant
6801 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6802 1 : .await?;
6803 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6804 :
6805 1 : tenant
6806 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6807 1 : .await?;
6808 1 : let newtline = tenant
6809 1 : .get_timeline(NEW_TIMELINE_ID, true)
6810 1 : .expect("Should have a local timeline");
6811 :
6812 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6813 :
6814 1 : tline.set_broken("test".to_owned());
6815 :
6816 1 : tenant
6817 1 : .gc_iteration(
6818 1 : Some(TIMELINE_ID),
6819 1 : 0x10,
6820 1 : Duration::ZERO,
6821 1 : &CancellationToken::new(),
6822 1 : &ctx,
6823 1 : )
6824 1 : .await?;
6825 :
6826 : // The branchpoints should contain all timelines, even ones marked
6827 : // as Broken.
6828 : {
6829 1 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
6830 1 : assert_eq!(branchpoints.len(), 1);
6831 1 : assert_eq!(
6832 1 : branchpoints[0],
6833 : (Lsn(0x40), NEW_TIMELINE_ID, MaybeOffloaded::No)
6834 : );
6835 : }
6836 :
6837 : // You can read the key from the child branch even though the parent is
6838 : // Broken, as long as you don't need to access data from the parent.
6839 1 : assert_eq!(
6840 1 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
6841 1 : test_img(&format!("foo at {}", Lsn(0x70)))
6842 : );
6843 :
6844 : // This needs to traverse to the parent, and fails.
6845 1 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
6846 1 : assert!(
6847 1 : err.to_string().starts_with(&format!(
6848 1 : "bad state on timeline {}: Broken",
6849 1 : tline.timeline_id
6850 1 : )),
6851 0 : "{err}"
6852 : );
6853 :
6854 2 : Ok(())
6855 1 : }
6856 :
6857 : #[tokio::test]
6858 1 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
6859 1 : let (tenant, ctx) =
6860 1 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
6861 1 : .await?
6862 1 : .load()
6863 1 : .await;
6864 1 : let tline = tenant
6865 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6866 1 : .await?;
6867 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6868 :
6869 1 : tenant
6870 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6871 1 : .await?;
6872 1 : let newtline = tenant
6873 1 : .get_timeline(NEW_TIMELINE_ID, true)
6874 1 : .expect("Should have a local timeline");
6875 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
6876 1 : tenant
6877 1 : .gc_iteration(
6878 1 : Some(TIMELINE_ID),
6879 1 : 0x10,
6880 1 : Duration::ZERO,
6881 1 : &CancellationToken::new(),
6882 1 : &ctx,
6883 1 : )
6884 1 : .await?;
6885 1 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
6886 :
6887 2 : Ok(())
6888 1 : }
6889 : #[tokio::test]
6890 1 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
6891 1 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
6892 1 : .await?
6893 1 : .load()
6894 1 : .await;
6895 1 : let tline = tenant
6896 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6897 1 : .await?;
6898 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6899 :
6900 1 : tenant
6901 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6902 1 : .await?;
6903 1 : let newtline = tenant
6904 1 : .get_timeline(NEW_TIMELINE_ID, true)
6905 1 : .expect("Should have a local timeline");
6906 :
6907 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6908 :
6909 : // run gc on parent
6910 1 : tenant
6911 1 : .gc_iteration(
6912 1 : Some(TIMELINE_ID),
6913 1 : 0x10,
6914 1 : Duration::ZERO,
6915 1 : &CancellationToken::new(),
6916 1 : &ctx,
6917 1 : )
6918 1 : .await?;
6919 :
6920 : // Check that the data is still accessible on the branch.
6921 1 : assert_eq!(
6922 1 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
6923 1 : test_img(&format!("foo at {}", Lsn(0x40)))
6924 : );
6925 :
6926 2 : Ok(())
6927 1 : }
6928 :
6929 : #[tokio::test]
6930 1 : async fn timeline_load() -> anyhow::Result<()> {
6931 : const TEST_NAME: &str = "timeline_load";
6932 1 : let harness = TenantHarness::create(TEST_NAME).await?;
6933 : {
6934 1 : let (tenant, ctx) = harness.load().await;
6935 1 : let tline = tenant
6936 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
6937 1 : .await?;
6938 1 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
6939 : // so that all uploads finish & we can call harness.load() below again
6940 1 : tenant
6941 1 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6942 1 : .instrument(harness.span())
6943 1 : .await
6944 1 : .ok()
6945 1 : .unwrap();
6946 : }
6947 :
6948 1 : let (tenant, _ctx) = harness.load().await;
6949 1 : tenant
6950 1 : .get_timeline(TIMELINE_ID, true)
6951 1 : .expect("cannot load timeline");
6952 :
6953 2 : Ok(())
6954 1 : }
6955 :
6956 : #[tokio::test]
6957 1 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
6958 : const TEST_NAME: &str = "timeline_load_with_ancestor";
6959 1 : let harness = TenantHarness::create(TEST_NAME).await?;
6960 : // create two timelines
6961 : {
6962 1 : let (tenant, ctx) = harness.load().await;
6963 1 : let tline = tenant
6964 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6965 1 : .await?;
6966 :
6967 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
6968 :
6969 1 : let child_tline = tenant
6970 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
6971 1 : .await?;
6972 1 : child_tline.set_state(TimelineState::Active);
6973 :
6974 1 : let newtline = tenant
6975 1 : .get_timeline(NEW_TIMELINE_ID, true)
6976 1 : .expect("Should have a local timeline");
6977 :
6978 1 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
6979 :
6980 : // so that all uploads finish & we can call harness.load() below again
6981 1 : tenant
6982 1 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
6983 1 : .instrument(harness.span())
6984 1 : .await
6985 1 : .ok()
6986 1 : .unwrap();
6987 : }
6988 :
6989 : // check that both of them are initially unloaded
6990 1 : let (tenant, _ctx) = harness.load().await;
6991 :
6992 : // check that both, child and ancestor are loaded
6993 1 : let _child_tline = tenant
6994 1 : .get_timeline(NEW_TIMELINE_ID, true)
6995 1 : .expect("cannot get child timeline loaded");
6996 :
6997 1 : let _ancestor_tline = tenant
6998 1 : .get_timeline(TIMELINE_ID, true)
6999 1 : .expect("cannot get ancestor timeline loaded");
7000 :
7001 2 : Ok(())
7002 1 : }
7003 :
7004 : #[tokio::test]
7005 1 : async fn delta_layer_dumping() -> anyhow::Result<()> {
7006 : use storage_layer::AsLayerDesc;
7007 1 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
7008 1 : .await?
7009 1 : .load()
7010 1 : .await;
7011 1 : let tline = tenant
7012 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7013 1 : .await?;
7014 1 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
7015 :
7016 1 : let layer_map = tline.layers.read(LayerManagerLockHolder::Testing).await;
7017 1 : let level0_deltas = layer_map
7018 1 : .layer_map()?
7019 1 : .level0_deltas()
7020 1 : .iter()
7021 2 : .map(|desc| layer_map.get_from_desc(desc))
7022 1 : .collect::<Vec<_>>();
7023 :
7024 1 : assert!(!level0_deltas.is_empty());
7025 :
7026 3 : for delta in level0_deltas {
7027 1 : // Ensure we are dumping a delta layer here
7028 2 : assert!(delta.layer_desc().is_delta);
7029 2 : delta.dump(true, &ctx).await.unwrap();
7030 1 : }
7031 1 :
7032 1 : Ok(())
7033 1 : }
7034 :
7035 : #[tokio::test]
7036 1 : async fn test_images() -> anyhow::Result<()> {
7037 1 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
7038 1 : let tline = tenant
7039 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7040 1 : .await?;
7041 :
7042 1 : let mut writer = tline.writer().await;
7043 1 : writer
7044 1 : .put(
7045 1 : *TEST_KEY,
7046 1 : Lsn(0x10),
7047 1 : &Value::Image(test_img("foo at 0x10")),
7048 1 : &ctx,
7049 1 : )
7050 1 : .await?;
7051 1 : writer.finish_write(Lsn(0x10));
7052 1 : drop(writer);
7053 :
7054 1 : tline.freeze_and_flush().await?;
7055 1 : tline
7056 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7057 1 : .await?;
7058 :
7059 1 : let mut writer = tline.writer().await;
7060 1 : writer
7061 1 : .put(
7062 1 : *TEST_KEY,
7063 1 : Lsn(0x20),
7064 1 : &Value::Image(test_img("foo at 0x20")),
7065 1 : &ctx,
7066 1 : )
7067 1 : .await?;
7068 1 : writer.finish_write(Lsn(0x20));
7069 1 : drop(writer);
7070 :
7071 1 : tline.freeze_and_flush().await?;
7072 1 : tline
7073 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7074 1 : .await?;
7075 :
7076 1 : let mut writer = tline.writer().await;
7077 1 : writer
7078 1 : .put(
7079 1 : *TEST_KEY,
7080 1 : Lsn(0x30),
7081 1 : &Value::Image(test_img("foo at 0x30")),
7082 1 : &ctx,
7083 1 : )
7084 1 : .await?;
7085 1 : writer.finish_write(Lsn(0x30));
7086 1 : drop(writer);
7087 :
7088 1 : tline.freeze_and_flush().await?;
7089 1 : tline
7090 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7091 1 : .await?;
7092 :
7093 1 : let mut writer = tline.writer().await;
7094 1 : writer
7095 1 : .put(
7096 1 : *TEST_KEY,
7097 1 : Lsn(0x40),
7098 1 : &Value::Image(test_img("foo at 0x40")),
7099 1 : &ctx,
7100 1 : )
7101 1 : .await?;
7102 1 : writer.finish_write(Lsn(0x40));
7103 1 : drop(writer);
7104 :
7105 1 : tline.freeze_and_flush().await?;
7106 1 : tline
7107 1 : .compact(&CancellationToken::new(), EnumSet::default(), &ctx)
7108 1 : .await?;
7109 :
7110 1 : assert_eq!(
7111 1 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
7112 1 : test_img("foo at 0x10")
7113 : );
7114 1 : assert_eq!(
7115 1 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
7116 1 : test_img("foo at 0x10")
7117 : );
7118 1 : assert_eq!(
7119 1 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
7120 1 : test_img("foo at 0x20")
7121 : );
7122 1 : assert_eq!(
7123 1 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
7124 1 : test_img("foo at 0x30")
7125 : );
7126 1 : assert_eq!(
7127 1 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
7128 1 : test_img("foo at 0x40")
7129 : );
7130 :
7131 2 : Ok(())
7132 1 : }
7133 :
7134 2 : async fn bulk_insert_compact_gc(
7135 2 : tenant: &TenantShard,
7136 2 : timeline: &Arc<Timeline>,
7137 2 : ctx: &RequestContext,
7138 2 : lsn: Lsn,
7139 2 : repeat: usize,
7140 2 : key_count: usize,
7141 2 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
7142 2 : let compact = true;
7143 2 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
7144 2 : }
7145 :
7146 4 : async fn bulk_insert_maybe_compact_gc(
7147 4 : tenant: &TenantShard,
7148 4 : timeline: &Arc<Timeline>,
7149 4 : ctx: &RequestContext,
7150 4 : mut lsn: Lsn,
7151 4 : repeat: usize,
7152 4 : key_count: usize,
7153 4 : compact: bool,
7154 4 : ) -> anyhow::Result<HashMap<Key, BTreeSet<Lsn>>> {
7155 4 : let mut inserted: HashMap<Key, BTreeSet<Lsn>> = Default::default();
7156 :
7157 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7158 4 : let mut blknum = 0;
7159 :
7160 : // Enforce that key range is monotonously increasing
7161 4 : let mut keyspace = KeySpaceAccum::new();
7162 :
7163 4 : let cancel = CancellationToken::new();
7164 :
7165 4 : for _ in 0..repeat {
7166 200 : for _ in 0..key_count {
7167 2000000 : test_key.field6 = blknum;
7168 2000000 : let mut writer = timeline.writer().await;
7169 2000000 : writer
7170 2000000 : .put(
7171 2000000 : test_key,
7172 2000000 : lsn,
7173 2000000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7174 2000000 : ctx,
7175 2000000 : )
7176 2000000 : .await?;
7177 2000000 : inserted.entry(test_key).or_default().insert(lsn);
7178 2000000 : writer.finish_write(lsn);
7179 2000000 : drop(writer);
7180 :
7181 2000000 : keyspace.add_key(test_key);
7182 :
7183 2000000 : lsn = Lsn(lsn.0 + 0x10);
7184 2000000 : blknum += 1;
7185 : }
7186 :
7187 200 : timeline.freeze_and_flush().await?;
7188 200 : if compact {
7189 : // this requires timeline to be &Arc<Timeline>
7190 100 : timeline.compact(&cancel, EnumSet::default(), ctx).await?;
7191 100 : }
7192 :
7193 : // this doesn't really need to use the timeline_id target, but it is closer to what it
7194 : // originally was.
7195 200 : let res = tenant
7196 200 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
7197 200 : .await?;
7198 :
7199 200 : assert_eq!(res.layers_removed, 0, "this never removes anything");
7200 : }
7201 :
7202 4 : Ok(inserted)
7203 4 : }
7204 :
7205 : //
7206 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
7207 : // Repeat 50 times.
7208 : //
7209 : #[tokio::test]
7210 1 : async fn test_bulk_insert() -> anyhow::Result<()> {
7211 1 : let harness = TenantHarness::create("test_bulk_insert").await?;
7212 1 : let (tenant, ctx) = harness.load().await;
7213 1 : let tline = tenant
7214 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7215 1 : .await?;
7216 :
7217 1 : let lsn = Lsn(0x10);
7218 1 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7219 :
7220 2 : Ok(())
7221 1 : }
7222 :
7223 : // Test the vectored get real implementation against a simple sequential implementation.
7224 : //
7225 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
7226 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
7227 : // grow to the right on the X axis.
7228 : // [Delta]
7229 : // [Delta]
7230 : // [Delta]
7231 : // [Delta]
7232 : // ------------ Image ---------------
7233 : //
7234 : // After layer generation we pick the ranges to query as follows:
7235 : // 1. The beginning of each delta layer
7236 : // 2. At the seam between two adjacent delta layers
7237 : //
7238 : // There's one major downside to this test: delta layers only contains images,
7239 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
7240 : #[tokio::test]
7241 1 : async fn test_get_vectored() -> anyhow::Result<()> {
7242 1 : let harness = TenantHarness::create("test_get_vectored").await?;
7243 1 : let (tenant, ctx) = harness.load().await;
7244 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7245 1 : let tline = tenant
7246 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
7247 1 : .await?;
7248 :
7249 1 : let lsn = Lsn(0x10);
7250 1 : let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
7251 :
7252 1 : let guard = tline.layers.read(LayerManagerLockHolder::Testing).await;
7253 1 : let lm = guard.layer_map()?;
7254 :
7255 1 : lm.dump(true, &ctx).await?;
7256 :
7257 1 : let mut reads = Vec::new();
7258 1 : let mut prev = None;
7259 6 : lm.iter_historic_layers().for_each(|desc| {
7260 6 : if !desc.is_delta() {
7261 1 : prev = Some(desc.clone());
7262 1 : return;
7263 5 : }
7264 :
7265 5 : let start = desc.key_range.start;
7266 5 : let end = desc
7267 5 : .key_range
7268 5 : .start
7269 5 : .add(tenant.conf.max_get_vectored_keys.get() as u32);
7270 5 : reads.push(KeySpace {
7271 5 : ranges: vec![start..end],
7272 5 : });
7273 :
7274 5 : if let Some(prev) = &prev {
7275 5 : if !prev.is_delta() {
7276 5 : return;
7277 0 : }
7278 :
7279 0 : let first_range = Key {
7280 0 : field6: prev.key_range.end.field6 - 4,
7281 0 : ..prev.key_range.end
7282 0 : }..prev.key_range.end;
7283 :
7284 0 : let second_range = desc.key_range.start..Key {
7285 0 : field6: desc.key_range.start.field6 + 4,
7286 0 : ..desc.key_range.start
7287 0 : };
7288 :
7289 0 : reads.push(KeySpace {
7290 0 : ranges: vec![first_range, second_range],
7291 0 : });
7292 0 : };
7293 :
7294 0 : prev = Some(desc.clone());
7295 6 : });
7296 :
7297 1 : drop(guard);
7298 :
7299 : // Pick a big LSN such that we query over all the changes.
7300 1 : let reads_lsn = Lsn(u64::MAX - 1);
7301 :
7302 6 : for read in reads {
7303 5 : info!("Doing vectored read on {:?}", read);
7304 1 :
7305 5 : let query = VersionedKeySpaceQuery::uniform(read.clone(), reads_lsn);
7306 1 :
7307 5 : let vectored_res = tline
7308 5 : .get_vectored_impl(
7309 5 : query,
7310 5 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7311 5 : &ctx,
7312 5 : )
7313 5 : .await;
7314 1 :
7315 5 : let mut expected_lsns: HashMap<Key, Lsn> = Default::default();
7316 5 : let mut expect_missing = false;
7317 5 : let mut key = read.start().unwrap();
7318 165 : while key != read.end().unwrap() {
7319 160 : if let Some(lsns) = inserted.get(&key) {
7320 160 : let expected_lsn = lsns.iter().rfind(|lsn| **lsn <= reads_lsn);
7321 160 : match expected_lsn {
7322 160 : Some(lsn) => {
7323 160 : expected_lsns.insert(key, *lsn);
7324 160 : }
7325 1 : None => {
7326 1 : expect_missing = true;
7327 1 : break;
7328 1 : }
7329 1 : }
7330 1 : } else {
7331 1 : expect_missing = true;
7332 1 : break;
7333 1 : }
7334 1 :
7335 160 : key = key.next();
7336 1 : }
7337 1 :
7338 5 : if expect_missing {
7339 1 : assert!(matches!(vectored_res, Err(GetVectoredError::MissingKey(_))));
7340 1 : } else {
7341 160 : for (key, image) in vectored_res? {
7342 160 : let expected_lsn = expected_lsns.get(&key).expect("determined above");
7343 160 : let expected_image = test_img(&format!("{} at {}", key.field6, expected_lsn));
7344 160 : assert_eq!(image?, expected_image);
7345 1 : }
7346 1 : }
7347 1 : }
7348 1 :
7349 1 : Ok(())
7350 1 : }
7351 :
7352 : #[tokio::test]
7353 1 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
7354 1 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
7355 :
7356 1 : let (tenant, ctx) = harness.load().await;
7357 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7358 1 : let (tline, ctx) = tenant
7359 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
7360 1 : .await?;
7361 1 : let tline = tline.raw_timeline().unwrap();
7362 :
7363 1 : let mut modification = tline.begin_modification(Lsn(0x1000));
7364 1 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
7365 1 : modification.set_lsn(Lsn(0x1008))?;
7366 1 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
7367 1 : modification.commit(&ctx).await?;
7368 :
7369 1 : let child_timeline_id = TimelineId::generate();
7370 1 : tenant
7371 1 : .branch_timeline_test(
7372 1 : tline,
7373 1 : child_timeline_id,
7374 1 : Some(tline.get_last_record_lsn()),
7375 1 : &ctx,
7376 1 : )
7377 1 : .await?;
7378 :
7379 1 : let child_timeline = tenant
7380 1 : .get_timeline(child_timeline_id, true)
7381 1 : .expect("Should have the branched timeline");
7382 :
7383 1 : let aux_keyspace = KeySpace {
7384 1 : ranges: vec![NON_INHERITED_RANGE],
7385 1 : };
7386 1 : let read_lsn = child_timeline.get_last_record_lsn();
7387 :
7388 1 : let query = VersionedKeySpaceQuery::uniform(aux_keyspace.clone(), read_lsn);
7389 :
7390 1 : let vectored_res = child_timeline
7391 1 : .get_vectored_impl(
7392 1 : query,
7393 1 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7394 1 : &ctx,
7395 1 : )
7396 1 : .await;
7397 :
7398 1 : let images = vectored_res?;
7399 1 : assert!(images.is_empty());
7400 2 : Ok(())
7401 1 : }
7402 :
7403 : // Test that vectored get handles layer gaps correctly
7404 : // by advancing into the next ancestor timeline if required.
7405 : //
7406 : // The test generates timelines that look like the diagram below.
7407 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
7408 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
7409 : //
7410 : // ```
7411 : //-------------------------------+
7412 : // ... |
7413 : // [ L1 ] |
7414 : // [ / L1 ] | Child Timeline
7415 : // ... |
7416 : // ------------------------------+
7417 : // [ X L1 ] | Parent Timeline
7418 : // ------------------------------+
7419 : // ```
7420 : #[tokio::test]
7421 1 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
7422 1 : let tenant_conf = pageserver_api::models::TenantConfig {
7423 1 : // Make compaction deterministic
7424 1 : gc_period: Some(Duration::ZERO),
7425 1 : compaction_period: Some(Duration::ZERO),
7426 1 : // Encourage creation of L1 layers
7427 1 : checkpoint_distance: Some(16 * 1024),
7428 1 : compaction_target_size: Some(8 * 1024),
7429 1 : ..Default::default()
7430 1 : };
7431 :
7432 1 : let harness = TenantHarness::create_custom(
7433 1 : "test_get_vectored_key_gap",
7434 1 : tenant_conf,
7435 1 : TenantId::generate(),
7436 1 : ShardIdentity::unsharded(),
7437 1 : Generation::new(0xdeadbeef),
7438 1 : )
7439 1 : .await?;
7440 1 : let (tenant, ctx) = harness.load().await;
7441 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7442 :
7443 1 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7444 1 : let gap_at_key = current_key.add(100);
7445 1 : let mut current_lsn = Lsn(0x10);
7446 :
7447 : const KEY_COUNT: usize = 10_000;
7448 :
7449 1 : let timeline_id = TimelineId::generate();
7450 1 : let current_timeline = tenant
7451 1 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7452 1 : .await?;
7453 :
7454 1 : current_lsn += 0x100;
7455 :
7456 1 : let mut writer = current_timeline.writer().await;
7457 1 : writer
7458 1 : .put(
7459 1 : gap_at_key,
7460 1 : current_lsn,
7461 1 : &Value::Image(test_img(&format!("{gap_at_key} at {current_lsn}"))),
7462 1 : &ctx,
7463 1 : )
7464 1 : .await?;
7465 1 : writer.finish_write(current_lsn);
7466 1 : drop(writer);
7467 :
7468 1 : let mut latest_lsns = HashMap::new();
7469 1 : latest_lsns.insert(gap_at_key, current_lsn);
7470 :
7471 1 : current_timeline.freeze_and_flush().await?;
7472 :
7473 1 : let child_timeline_id = TimelineId::generate();
7474 :
7475 1 : tenant
7476 1 : .branch_timeline_test(
7477 1 : ¤t_timeline,
7478 1 : child_timeline_id,
7479 1 : Some(current_lsn),
7480 1 : &ctx,
7481 1 : )
7482 1 : .await?;
7483 1 : let child_timeline = tenant
7484 1 : .get_timeline(child_timeline_id, true)
7485 1 : .expect("Should have the branched timeline");
7486 :
7487 10001 : for i in 0..KEY_COUNT {
7488 10000 : if current_key == gap_at_key {
7489 1 : current_key = current_key.next();
7490 1 : continue;
7491 9999 : }
7492 :
7493 9999 : current_lsn += 0x10;
7494 :
7495 9999 : let mut writer = child_timeline.writer().await;
7496 9999 : writer
7497 9999 : .put(
7498 9999 : current_key,
7499 9999 : current_lsn,
7500 9999 : &Value::Image(test_img(&format!("{current_key} at {current_lsn}"))),
7501 9999 : &ctx,
7502 9999 : )
7503 9999 : .await?;
7504 9999 : writer.finish_write(current_lsn);
7505 9999 : drop(writer);
7506 :
7507 9999 : latest_lsns.insert(current_key, current_lsn);
7508 9999 : current_key = current_key.next();
7509 :
7510 : // Flush every now and then to encourage layer file creation.
7511 9999 : if i % 500 == 0 {
7512 20 : child_timeline.freeze_and_flush().await?;
7513 9979 : }
7514 : }
7515 :
7516 1 : child_timeline.freeze_and_flush().await?;
7517 1 : let mut flags = EnumSet::new();
7518 1 : flags.insert(CompactFlags::ForceRepartition);
7519 1 : child_timeline
7520 1 : .compact(&CancellationToken::new(), flags, &ctx)
7521 1 : .await?;
7522 :
7523 1 : let key_near_end = {
7524 1 : let mut tmp = current_key;
7525 1 : tmp.field6 -= 10;
7526 1 : tmp
7527 : };
7528 :
7529 1 : let key_near_gap = {
7530 1 : let mut tmp = gap_at_key;
7531 1 : tmp.field6 -= 10;
7532 1 : tmp
7533 : };
7534 :
7535 1 : let read = KeySpace {
7536 1 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
7537 1 : };
7538 :
7539 1 : let query = VersionedKeySpaceQuery::uniform(read.clone(), current_lsn);
7540 :
7541 1 : let results = child_timeline
7542 1 : .get_vectored_impl(
7543 1 : query,
7544 1 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7545 1 : &ctx,
7546 1 : )
7547 1 : .await?;
7548 :
7549 22 : for (key, img_res) in results {
7550 21 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
7551 21 : assert_eq!(img_res?, expected);
7552 1 : }
7553 1 :
7554 1 : Ok(())
7555 1 : }
7556 :
7557 : // Test that vectored get descends into ancestor timelines correctly and
7558 : // does not return an image that's newer than requested.
7559 : //
7560 : // The diagram below ilustrates an interesting case. We have a parent timeline
7561 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
7562 : // from the child timeline, so the parent timeline must be visited. When advacing into
7563 : // the child timeline, the read path needs to remember what the requested Lsn was in
7564 : // order to avoid returning an image that's too new. The test below constructs such
7565 : // a timeline setup and does a few queries around the Lsn of each page image.
7566 : // ```
7567 : // LSN
7568 : // ^
7569 : // |
7570 : // |
7571 : // 500 | --------------------------------------> branch point
7572 : // 400 | X
7573 : // 300 | X
7574 : // 200 | --------------------------------------> requested lsn
7575 : // 100 | X
7576 : // |---------------------------------------> Key
7577 : // |
7578 : // ------> requested key
7579 : //
7580 : // Legend:
7581 : // * X - page images
7582 : // ```
7583 : #[tokio::test]
7584 1 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
7585 1 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
7586 1 : let (tenant, ctx) = harness.load().await;
7587 1 : let io_concurrency = IoConcurrency::spawn_for_test();
7588 :
7589 1 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7590 1 : let end_key = start_key.add(1000);
7591 1 : let child_gap_at_key = start_key.add(500);
7592 1 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
7593 :
7594 1 : let mut current_lsn = Lsn(0x10);
7595 :
7596 1 : let timeline_id = TimelineId::generate();
7597 1 : let parent_timeline = tenant
7598 1 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
7599 1 : .await?;
7600 :
7601 1 : current_lsn += 0x100;
7602 :
7603 4 : for _ in 0..3 {
7604 3 : let mut key = start_key;
7605 3003 : while key < end_key {
7606 3000 : current_lsn += 0x10;
7607 :
7608 3000 : let image_value = format!("{child_gap_at_key} at {current_lsn}");
7609 :
7610 3000 : let mut writer = parent_timeline.writer().await;
7611 3000 : writer
7612 3000 : .put(
7613 3000 : key,
7614 3000 : current_lsn,
7615 3000 : &Value::Image(test_img(&image_value)),
7616 3000 : &ctx,
7617 3000 : )
7618 3000 : .await?;
7619 3000 : writer.finish_write(current_lsn);
7620 :
7621 3000 : if key == child_gap_at_key {
7622 3 : parent_gap_lsns.insert(current_lsn, image_value);
7623 2997 : }
7624 :
7625 3000 : key = key.next();
7626 : }
7627 :
7628 3 : parent_timeline.freeze_and_flush().await?;
7629 : }
7630 :
7631 1 : let child_timeline_id = TimelineId::generate();
7632 :
7633 1 : let child_timeline = tenant
7634 1 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
7635 1 : .await?;
7636 :
7637 1 : let mut key = start_key;
7638 1001 : while key < end_key {
7639 1000 : if key == child_gap_at_key {
7640 1 : key = key.next();
7641 1 : continue;
7642 999 : }
7643 :
7644 999 : current_lsn += 0x10;
7645 :
7646 999 : let mut writer = child_timeline.writer().await;
7647 999 : writer
7648 999 : .put(
7649 999 : key,
7650 999 : current_lsn,
7651 999 : &Value::Image(test_img(&format!("{key} at {current_lsn}"))),
7652 999 : &ctx,
7653 999 : )
7654 999 : .await?;
7655 999 : writer.finish_write(current_lsn);
7656 :
7657 999 : key = key.next();
7658 : }
7659 :
7660 1 : child_timeline.freeze_and_flush().await?;
7661 :
7662 1 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
7663 1 : let mut query_lsns = Vec::new();
7664 3 : for image_lsn in parent_gap_lsns.keys().rev() {
7665 18 : for offset in lsn_offsets {
7666 15 : query_lsns.push(Lsn(image_lsn
7667 15 : .0
7668 15 : .checked_add_signed(offset)
7669 15 : .expect("Shouldn't overflow")));
7670 15 : }
7671 1 : }
7672 1 :
7673 16 : for query_lsn in query_lsns {
7674 15 : let query = VersionedKeySpaceQuery::uniform(
7675 15 : KeySpace {
7676 15 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
7677 15 : },
7678 15 : query_lsn,
7679 1 : );
7680 1 :
7681 15 : let results = child_timeline
7682 15 : .get_vectored_impl(
7683 15 : query,
7684 15 : &mut ValuesReconstructState::new(io_concurrency.clone()),
7685 15 : &ctx,
7686 15 : )
7687 15 : .await;
7688 1 :
7689 15 : let expected_item = parent_gap_lsns
7690 15 : .iter()
7691 15 : .rev()
7692 34 : .find(|(lsn, _)| **lsn <= query_lsn);
7693 1 :
7694 15 : info!(
7695 1 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
7696 1 : query_lsn, expected_item
7697 1 : );
7698 1 :
7699 15 : match expected_item {
7700 13 : Some((_, img_value)) => {
7701 13 : let key_results = results.expect("No vectored get error expected");
7702 13 : let key_result = &key_results[&child_gap_at_key];
7703 13 : let returned_img = key_result
7704 13 : .as_ref()
7705 13 : .expect("No page reconstruct error expected");
7706 1 :
7707 13 : info!(
7708 1 : "Vectored read at LSN {} returned image {}",
7709 1 : query_lsn,
7710 1 : std::str::from_utf8(returned_img)?
7711 1 : );
7712 13 : assert_eq!(*returned_img, test_img(img_value));
7713 1 : }
7714 1 : None => {
7715 2 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
7716 1 : }
7717 1 : }
7718 1 : }
7719 1 :
7720 1 : Ok(())
7721 1 : }
7722 :
7723 : #[tokio::test]
7724 1 : async fn test_random_updates() -> anyhow::Result<()> {
7725 1 : let names_algorithms = [
7726 1 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
7727 1 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
7728 1 : ];
7729 3 : for (name, algorithm) in names_algorithms {
7730 2 : test_random_updates_algorithm(name, algorithm).await?;
7731 1 : }
7732 1 : Ok(())
7733 1 : }
7734 :
7735 2 : async fn test_random_updates_algorithm(
7736 2 : name: &'static str,
7737 2 : compaction_algorithm: CompactionAlgorithm,
7738 2 : ) -> anyhow::Result<()> {
7739 2 : let mut harness = TenantHarness::create(name).await?;
7740 2 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
7741 2 : kind: compaction_algorithm,
7742 2 : });
7743 2 : let (tenant, ctx) = harness.load().await;
7744 2 : let tline = tenant
7745 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7746 2 : .await?;
7747 :
7748 : const NUM_KEYS: usize = 1000;
7749 2 : let cancel = CancellationToken::new();
7750 :
7751 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7752 2 : let mut test_key_end = test_key;
7753 2 : test_key_end.field6 = NUM_KEYS as u32;
7754 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
7755 :
7756 2 : let mut keyspace = KeySpaceAccum::new();
7757 :
7758 : // Track when each page was last modified. Used to assert that
7759 : // a read sees the latest page version.
7760 2 : let mut updated = [Lsn(0); NUM_KEYS];
7761 :
7762 2 : let mut lsn = Lsn(0x10);
7763 : #[allow(clippy::needless_range_loop)]
7764 2002 : for blknum in 0..NUM_KEYS {
7765 2000 : lsn = Lsn(lsn.0 + 0x10);
7766 2000 : test_key.field6 = blknum as u32;
7767 2000 : let mut writer = tline.writer().await;
7768 2000 : writer
7769 2000 : .put(
7770 2000 : test_key,
7771 2000 : lsn,
7772 2000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7773 2000 : &ctx,
7774 2000 : )
7775 2000 : .await?;
7776 2000 : writer.finish_write(lsn);
7777 2000 : updated[blknum] = lsn;
7778 2000 : drop(writer);
7779 :
7780 2000 : keyspace.add_key(test_key);
7781 : }
7782 :
7783 102 : for _ in 0..50 {
7784 100100 : for _ in 0..NUM_KEYS {
7785 100000 : lsn = Lsn(lsn.0 + 0x10);
7786 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7787 100000 : test_key.field6 = blknum as u32;
7788 100000 : let mut writer = tline.writer().await;
7789 100000 : writer
7790 100000 : .put(
7791 100000 : test_key,
7792 100000 : lsn,
7793 100000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7794 100000 : &ctx,
7795 100000 : )
7796 100000 : .await?;
7797 100000 : writer.finish_write(lsn);
7798 100000 : drop(writer);
7799 100000 : updated[blknum] = lsn;
7800 : }
7801 :
7802 : // Read all the blocks
7803 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7804 100000 : test_key.field6 = blknum as u32;
7805 100000 : assert_eq!(
7806 100000 : tline.get(test_key, lsn, &ctx).await?,
7807 100000 : test_img(&format!("{blknum} at {last_lsn}"))
7808 : );
7809 : }
7810 :
7811 : // Perform a cycle of flush, and GC
7812 100 : tline.freeze_and_flush().await?;
7813 100 : tenant
7814 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7815 100 : .await?;
7816 : }
7817 :
7818 2 : Ok(())
7819 2 : }
7820 :
7821 : #[tokio::test]
7822 1 : async fn test_traverse_branches() -> anyhow::Result<()> {
7823 1 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
7824 1 : .await?
7825 1 : .load()
7826 1 : .await;
7827 1 : let mut tline = tenant
7828 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7829 1 : .await?;
7830 :
7831 : const NUM_KEYS: usize = 1000;
7832 :
7833 1 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7834 :
7835 1 : let mut keyspace = KeySpaceAccum::new();
7836 :
7837 1 : let cancel = CancellationToken::new();
7838 :
7839 : // Track when each page was last modified. Used to assert that
7840 : // a read sees the latest page version.
7841 1 : let mut updated = [Lsn(0); NUM_KEYS];
7842 :
7843 1 : let mut lsn = Lsn(0x10);
7844 1 : #[allow(clippy::needless_range_loop)]
7845 1001 : for blknum in 0..NUM_KEYS {
7846 1000 : lsn = Lsn(lsn.0 + 0x10);
7847 1000 : test_key.field6 = blknum as u32;
7848 1000 : let mut writer = tline.writer().await;
7849 1000 : writer
7850 1000 : .put(
7851 1000 : test_key,
7852 1000 : lsn,
7853 1000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7854 1000 : &ctx,
7855 1000 : )
7856 1000 : .await?;
7857 1000 : writer.finish_write(lsn);
7858 1000 : updated[blknum] = lsn;
7859 1000 : drop(writer);
7860 1 :
7861 1000 : keyspace.add_key(test_key);
7862 1 : }
7863 1 :
7864 51 : for _ in 0..50 {
7865 50 : let new_tline_id = TimelineId::generate();
7866 50 : tenant
7867 50 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7868 50 : .await?;
7869 50 : tline = tenant
7870 50 : .get_timeline(new_tline_id, true)
7871 50 : .expect("Should have the branched timeline");
7872 1 :
7873 50050 : for _ in 0..NUM_KEYS {
7874 50000 : lsn = Lsn(lsn.0 + 0x10);
7875 50000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7876 50000 : test_key.field6 = blknum as u32;
7877 50000 : let mut writer = tline.writer().await;
7878 50000 : writer
7879 50000 : .put(
7880 50000 : test_key,
7881 50000 : lsn,
7882 50000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
7883 50000 : &ctx,
7884 50000 : )
7885 50000 : .await?;
7886 50000 : println!("updating {blknum} at {lsn}");
7887 50000 : writer.finish_write(lsn);
7888 50000 : drop(writer);
7889 50000 : updated[blknum] = lsn;
7890 1 : }
7891 1 :
7892 1 : // Read all the blocks
7893 50000 : for (blknum, last_lsn) in updated.iter().enumerate() {
7894 50000 : test_key.field6 = blknum as u32;
7895 50000 : assert_eq!(
7896 50000 : tline.get(test_key, lsn, &ctx).await?,
7897 50000 : test_img(&format!("{blknum} at {last_lsn}"))
7898 1 : );
7899 1 : }
7900 1 :
7901 1 : // Perform a cycle of flush, compact, and GC
7902 50 : tline.freeze_and_flush().await?;
7903 50 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
7904 50 : tenant
7905 50 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
7906 50 : .await?;
7907 1 : }
7908 1 :
7909 1 : Ok(())
7910 1 : }
7911 :
7912 : #[tokio::test]
7913 1 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
7914 1 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
7915 1 : .await?
7916 1 : .load()
7917 1 : .await;
7918 1 : let mut tline = tenant
7919 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
7920 1 : .await?;
7921 :
7922 : const NUM_KEYS: usize = 100;
7923 : const NUM_TLINES: usize = 50;
7924 :
7925 1 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
7926 : // Track page mutation lsns across different timelines.
7927 1 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
7928 :
7929 1 : let mut lsn = Lsn(0x10);
7930 :
7931 1 : #[allow(clippy::needless_range_loop)]
7932 51 : for idx in 0..NUM_TLINES {
7933 50 : let new_tline_id = TimelineId::generate();
7934 50 : tenant
7935 50 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
7936 50 : .await?;
7937 50 : tline = tenant
7938 50 : .get_timeline(new_tline_id, true)
7939 50 : .expect("Should have the branched timeline");
7940 1 :
7941 5050 : for _ in 0..NUM_KEYS {
7942 5000 : lsn = Lsn(lsn.0 + 0x10);
7943 5000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
7944 5000 : test_key.field6 = blknum as u32;
7945 5000 : let mut writer = tline.writer().await;
7946 5000 : writer
7947 5000 : .put(
7948 5000 : test_key,
7949 5000 : lsn,
7950 5000 : &Value::Image(test_img(&format!("{idx} {blknum} at {lsn}"))),
7951 5000 : &ctx,
7952 5000 : )
7953 5000 : .await?;
7954 5000 : println!("updating [{idx}][{blknum}] at {lsn}");
7955 5000 : writer.finish_write(lsn);
7956 5000 : drop(writer);
7957 5000 : updated[idx][blknum] = lsn;
7958 1 : }
7959 1 : }
7960 1 :
7961 1 : // Read pages from leaf timeline across all ancestors.
7962 50 : for (idx, lsns) in updated.iter().enumerate() {
7963 5000 : for (blknum, lsn) in lsns.iter().enumerate() {
7964 1 : // Skip empty mutations.
7965 5000 : if lsn.0 == 0 {
7966 1847 : continue;
7967 3153 : }
7968 3153 : println!("checking [{idx}][{blknum}] at {lsn}");
7969 3153 : test_key.field6 = blknum as u32;
7970 3153 : assert_eq!(
7971 3153 : tline.get(test_key, *lsn, &ctx).await?,
7972 3153 : test_img(&format!("{idx} {blknum} at {lsn}"))
7973 1 : );
7974 1 : }
7975 1 : }
7976 1 : Ok(())
7977 1 : }
7978 :
7979 : #[tokio::test]
7980 1 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
7981 1 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
7982 1 : .await?
7983 1 : .load()
7984 1 : .await;
7985 :
7986 1 : let initdb_lsn = Lsn(0x20);
7987 1 : let (utline, ctx) = tenant
7988 1 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
7989 1 : .await?;
7990 1 : let tline = utline.raw_timeline().unwrap();
7991 :
7992 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
7993 1 : tline.maybe_spawn_flush_loop();
7994 :
7995 : // Make sure the timeline has the minimum set of required keys for operation.
7996 : // The only operation you can always do on an empty timeline is to `put` new data.
7997 : // Except if you `put` at `initdb_lsn`.
7998 : // In that case, there's an optimization to directly create image layers instead of delta layers.
7999 : // It uses `repartition()`, which assumes some keys to be present.
8000 : // Let's make sure the test timeline can handle that case.
8001 : {
8002 1 : let mut state = tline.flush_loop_state.lock().unwrap();
8003 1 : assert_eq!(
8004 : timeline::FlushLoopState::Running {
8005 : expect_initdb_optimization: false,
8006 : initdb_optimization_count: 0,
8007 : },
8008 1 : *state
8009 : );
8010 1 : *state = timeline::FlushLoopState::Running {
8011 1 : expect_initdb_optimization: true,
8012 1 : initdb_optimization_count: 0,
8013 1 : };
8014 : }
8015 :
8016 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
8017 : // As explained above, the optimization requires some keys to be present.
8018 : // As per `create_empty_timeline` documentation, use init_empty to set them.
8019 : // This is what `create_test_timeline` does, by the way.
8020 1 : let mut modification = tline.begin_modification(initdb_lsn);
8021 1 : modification
8022 1 : .init_empty_test_timeline()
8023 1 : .context("init_empty_test_timeline")?;
8024 1 : modification
8025 1 : .commit(&ctx)
8026 1 : .await
8027 1 : .context("commit init_empty_test_timeline modification")?;
8028 :
8029 : // Do the flush. The flush code will check the expectations that we set above.
8030 1 : tline.freeze_and_flush().await?;
8031 :
8032 : // assert freeze_and_flush exercised the initdb optimization
8033 1 : {
8034 1 : let state = tline.flush_loop_state.lock().unwrap();
8035 1 : let timeline::FlushLoopState::Running {
8036 1 : expect_initdb_optimization,
8037 1 : initdb_optimization_count,
8038 1 : } = *state
8039 1 : else {
8040 1 : panic!("unexpected state: {:?}", *state);
8041 1 : };
8042 1 : assert!(expect_initdb_optimization);
8043 1 : assert!(initdb_optimization_count > 0);
8044 1 : }
8045 1 : Ok(())
8046 1 : }
8047 :
8048 : #[tokio::test]
8049 1 : async fn test_create_guard_crash() -> anyhow::Result<()> {
8050 1 : let name = "test_create_guard_crash";
8051 1 : let harness = TenantHarness::create(name).await?;
8052 : {
8053 1 : let (tenant, ctx) = harness.load().await;
8054 1 : let (tline, _ctx) = tenant
8055 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
8056 1 : .await?;
8057 : // Leave the timeline ID in [`TenantShard::timelines_creating`] to exclude attempting to create it again
8058 1 : let raw_tline = tline.raw_timeline().unwrap();
8059 1 : raw_tline
8060 1 : .shutdown(super::timeline::ShutdownMode::Hard)
8061 1 : .instrument(info_span!("test_shutdown", tenant_id=%raw_tline.tenant_shard_id, shard_id=%raw_tline.tenant_shard_id.shard_slug(), timeline_id=%TIMELINE_ID))
8062 1 : .await;
8063 1 : std::mem::forget(tline);
8064 : }
8065 :
8066 1 : let (tenant, _) = harness.load().await;
8067 1 : match tenant.get_timeline(TIMELINE_ID, false) {
8068 0 : Ok(_) => panic!("timeline should've been removed during load"),
8069 1 : Err(e) => {
8070 1 : assert_eq!(
8071 : e,
8072 1 : GetTimelineError::NotFound {
8073 1 : tenant_id: tenant.tenant_shard_id,
8074 1 : timeline_id: TIMELINE_ID,
8075 1 : }
8076 : )
8077 : }
8078 : }
8079 :
8080 1 : assert!(
8081 1 : !harness
8082 1 : .conf
8083 1 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
8084 1 : .exists()
8085 : );
8086 :
8087 2 : Ok(())
8088 1 : }
8089 :
8090 : #[tokio::test]
8091 1 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
8092 1 : let names_algorithms = [
8093 1 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
8094 1 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
8095 1 : ];
8096 3 : for (name, algorithm) in names_algorithms {
8097 2 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
8098 1 : }
8099 1 : Ok(())
8100 1 : }
8101 :
8102 2 : async fn test_read_at_max_lsn_algorithm(
8103 2 : name: &'static str,
8104 2 : compaction_algorithm: CompactionAlgorithm,
8105 2 : ) -> anyhow::Result<()> {
8106 2 : let mut harness = TenantHarness::create(name).await?;
8107 2 : harness.tenant_conf.compaction_algorithm = Some(CompactionAlgorithmSettings {
8108 2 : kind: compaction_algorithm,
8109 2 : });
8110 2 : let (tenant, ctx) = harness.load().await;
8111 2 : let tline = tenant
8112 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
8113 2 : .await?;
8114 :
8115 2 : let lsn = Lsn(0x10);
8116 2 : let compact = false;
8117 2 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
8118 :
8119 2 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
8120 2 : let read_lsn = Lsn(u64::MAX - 1);
8121 :
8122 2 : let result = tline.get(test_key, read_lsn, &ctx).await;
8123 2 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
8124 :
8125 2 : Ok(())
8126 2 : }
8127 :
8128 : #[tokio::test]
8129 1 : async fn test_metadata_scan() -> anyhow::Result<()> {
8130 1 : let harness = TenantHarness::create("test_metadata_scan").await?;
8131 1 : let (tenant, ctx) = harness.load().await;
8132 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8133 1 : let tline = tenant
8134 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8135 1 : .await?;
8136 :
8137 : const NUM_KEYS: usize = 1000;
8138 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
8139 :
8140 1 : let cancel = CancellationToken::new();
8141 :
8142 1 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8143 1 : base_key.field1 = AUX_KEY_PREFIX;
8144 1 : let mut test_key = base_key;
8145 :
8146 : // Track when each page was last modified. Used to assert that
8147 : // a read sees the latest page version.
8148 1 : let mut updated = [Lsn(0); NUM_KEYS];
8149 :
8150 1 : let mut lsn = Lsn(0x10);
8151 : #[allow(clippy::needless_range_loop)]
8152 1001 : for blknum in 0..NUM_KEYS {
8153 1000 : lsn = Lsn(lsn.0 + 0x10);
8154 1000 : test_key.field6 = (blknum * STEP) as u32;
8155 1000 : let mut writer = tline.writer().await;
8156 1000 : writer
8157 1000 : .put(
8158 1000 : test_key,
8159 1000 : lsn,
8160 1000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8161 1000 : &ctx,
8162 1000 : )
8163 1000 : .await?;
8164 1000 : writer.finish_write(lsn);
8165 1000 : updated[blknum] = lsn;
8166 1000 : drop(writer);
8167 : }
8168 :
8169 1 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
8170 :
8171 12 : for iter in 0..=10 {
8172 1 : // Read all the blocks
8173 11000 : for (blknum, last_lsn) in updated.iter().enumerate() {
8174 11000 : test_key.field6 = (blknum * STEP) as u32;
8175 11000 : assert_eq!(
8176 11000 : tline.get(test_key, lsn, &ctx).await?,
8177 11000 : test_img(&format!("{blknum} at {last_lsn}"))
8178 1 : );
8179 1 : }
8180 1 :
8181 11 : let mut cnt = 0;
8182 11 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8183 1 :
8184 11000 : for (key, value) in tline
8185 11 : .get_vectored_impl(
8186 11 : query,
8187 11 : &mut ValuesReconstructState::new(io_concurrency.clone()),
8188 11 : &ctx,
8189 11 : )
8190 11 : .await?
8191 1 : {
8192 11000 : let blknum = key.field6 as usize;
8193 11000 : let value = value?;
8194 11000 : assert!(blknum % STEP == 0);
8195 11000 : let blknum = blknum / STEP;
8196 11000 : assert_eq!(
8197 1 : value,
8198 11000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
8199 1 : );
8200 11000 : cnt += 1;
8201 1 : }
8202 1 :
8203 11 : assert_eq!(cnt, NUM_KEYS);
8204 1 :
8205 11011 : for _ in 0..NUM_KEYS {
8206 11000 : lsn = Lsn(lsn.0 + 0x10);
8207 11000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
8208 11000 : test_key.field6 = (blknum * STEP) as u32;
8209 11000 : let mut writer = tline.writer().await;
8210 11000 : writer
8211 11000 : .put(
8212 11000 : test_key,
8213 11000 : lsn,
8214 11000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8215 11000 : &ctx,
8216 11000 : )
8217 11000 : .await?;
8218 11000 : writer.finish_write(lsn);
8219 11000 : drop(writer);
8220 11000 : updated[blknum] = lsn;
8221 1 : }
8222 1 :
8223 1 : // Perform two cycles of flush, compact, and GC
8224 33 : for round in 0..2 {
8225 22 : tline.freeze_and_flush().await?;
8226 22 : tline
8227 22 : .compact(
8228 22 : &cancel,
8229 22 : if iter % 5 == 0 && round == 0 {
8230 3 : let mut flags = EnumSet::new();
8231 3 : flags.insert(CompactFlags::ForceImageLayerCreation);
8232 3 : flags.insert(CompactFlags::ForceRepartition);
8233 3 : flags
8234 1 : } else {
8235 19 : EnumSet::empty()
8236 1 : },
8237 22 : &ctx,
8238 1 : )
8239 22 : .await?;
8240 22 : tenant
8241 22 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
8242 22 : .await?;
8243 1 : }
8244 1 : }
8245 1 :
8246 1 : Ok(())
8247 1 : }
8248 :
8249 : #[tokio::test]
8250 1 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
8251 1 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
8252 1 : let (tenant, ctx) = harness.load().await;
8253 1 : let tline = tenant
8254 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8255 1 : .await?;
8256 :
8257 1 : let cancel = CancellationToken::new();
8258 :
8259 1 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8260 1 : base_key.field1 = AUX_KEY_PREFIX;
8261 1 : let test_key = base_key;
8262 1 : let mut lsn = Lsn(0x10);
8263 :
8264 21 : for _ in 0..20 {
8265 20 : lsn = Lsn(lsn.0 + 0x10);
8266 20 : let mut writer = tline.writer().await;
8267 20 : writer
8268 20 : .put(
8269 20 : test_key,
8270 20 : lsn,
8271 20 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
8272 20 : &ctx,
8273 20 : )
8274 20 : .await?;
8275 20 : writer.finish_write(lsn);
8276 20 : drop(writer);
8277 20 : tline.freeze_and_flush().await?; // force create a delta layer
8278 : }
8279 :
8280 1 : let before_num_l0_delta_files = tline
8281 1 : .layers
8282 1 : .read(LayerManagerLockHolder::Testing)
8283 1 : .await
8284 1 : .layer_map()?
8285 1 : .level0_deltas()
8286 1 : .len();
8287 :
8288 1 : tline.compact(&cancel, EnumSet::default(), &ctx).await?;
8289 :
8290 1 : let after_num_l0_delta_files = tline
8291 1 : .layers
8292 1 : .read(LayerManagerLockHolder::Testing)
8293 1 : .await
8294 1 : .layer_map()?
8295 1 : .level0_deltas()
8296 1 : .len();
8297 :
8298 1 : assert!(
8299 1 : after_num_l0_delta_files < before_num_l0_delta_files,
8300 0 : "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}"
8301 : );
8302 :
8303 1 : assert_eq!(
8304 1 : tline.get(test_key, lsn, &ctx).await?,
8305 1 : test_img(&format!("{} at {}", 0, lsn))
8306 : );
8307 :
8308 2 : Ok(())
8309 1 : }
8310 :
8311 : #[tokio::test]
8312 1 : async fn test_aux_file_e2e() {
8313 1 : let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
8314 :
8315 1 : let (tenant, ctx) = harness.load().await;
8316 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8317 :
8318 1 : let mut lsn = Lsn(0x08);
8319 :
8320 1 : let tline: Arc<Timeline> = tenant
8321 1 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8322 1 : .await
8323 1 : .unwrap();
8324 :
8325 : {
8326 1 : lsn += 8;
8327 1 : let mut modification = tline.begin_modification(lsn);
8328 1 : modification
8329 1 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
8330 1 : .await
8331 1 : .unwrap();
8332 1 : modification.commit(&ctx).await.unwrap();
8333 : }
8334 :
8335 : // we can read everything from the storage
8336 1 : let files = tline
8337 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8338 1 : .await
8339 1 : .unwrap();
8340 1 : assert_eq!(
8341 1 : files.get("pg_logical/mappings/test1"),
8342 1 : Some(&bytes::Bytes::from_static(b"first"))
8343 : );
8344 :
8345 : {
8346 1 : lsn += 8;
8347 1 : let mut modification = tline.begin_modification(lsn);
8348 1 : modification
8349 1 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
8350 1 : .await
8351 1 : .unwrap();
8352 1 : modification.commit(&ctx).await.unwrap();
8353 : }
8354 :
8355 1 : let files = tline
8356 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8357 1 : .await
8358 1 : .unwrap();
8359 1 : assert_eq!(
8360 1 : files.get("pg_logical/mappings/test2"),
8361 1 : Some(&bytes::Bytes::from_static(b"second"))
8362 : );
8363 :
8364 1 : let child = tenant
8365 1 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
8366 1 : .await
8367 1 : .unwrap();
8368 :
8369 1 : let files = child
8370 1 : .list_aux_files(lsn, &ctx, io_concurrency.clone())
8371 1 : .await
8372 1 : .unwrap();
8373 1 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
8374 1 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
8375 1 : }
8376 :
8377 : #[tokio::test]
8378 1 : async fn test_repl_origin_tombstones() {
8379 1 : let harness = TenantHarness::create("test_repl_origin_tombstones")
8380 1 : .await
8381 1 : .unwrap();
8382 :
8383 1 : let (tenant, ctx) = harness.load().await;
8384 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8385 :
8386 1 : let mut lsn = Lsn(0x08);
8387 :
8388 1 : let tline: Arc<Timeline> = tenant
8389 1 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
8390 1 : .await
8391 1 : .unwrap();
8392 :
8393 1 : let repl_lsn = Lsn(0x10);
8394 : {
8395 1 : lsn += 8;
8396 1 : let mut modification = tline.begin_modification(lsn);
8397 1 : modification.put_for_unit_test(repl_origin_key(2), Value::Image(Bytes::new()));
8398 1 : modification.set_replorigin(1, repl_lsn).await.unwrap();
8399 1 : modification.commit(&ctx).await.unwrap();
8400 : }
8401 :
8402 : // we can read everything from the storage
8403 1 : let repl_origins = tline
8404 1 : .get_replorigins(lsn, &ctx, io_concurrency.clone())
8405 1 : .await
8406 1 : .unwrap();
8407 1 : assert_eq!(repl_origins.len(), 1);
8408 1 : assert_eq!(repl_origins[&1], lsn);
8409 :
8410 : {
8411 1 : lsn += 8;
8412 1 : let mut modification = tline.begin_modification(lsn);
8413 1 : modification.put_for_unit_test(
8414 1 : repl_origin_key(3),
8415 1 : Value::Image(Bytes::copy_from_slice(b"cannot_decode_this")),
8416 : );
8417 1 : modification.commit(&ctx).await.unwrap();
8418 : }
8419 1 : let result = tline
8420 1 : .get_replorigins(lsn, &ctx, io_concurrency.clone())
8421 1 : .await;
8422 1 : assert!(result.is_err());
8423 1 : }
8424 :
8425 : #[tokio::test]
8426 1 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
8427 1 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
8428 1 : let (tenant, ctx) = harness.load().await;
8429 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8430 1 : let tline = tenant
8431 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
8432 1 : .await?;
8433 :
8434 : const NUM_KEYS: usize = 1000;
8435 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
8436 :
8437 1 : let cancel = CancellationToken::new();
8438 :
8439 1 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8440 1 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8441 1 : let mut test_key = base_key;
8442 1 : let mut lsn = Lsn(0x10);
8443 :
8444 4 : async fn scan_with_statistics(
8445 4 : tline: &Timeline,
8446 4 : keyspace: &KeySpace,
8447 4 : lsn: Lsn,
8448 4 : ctx: &RequestContext,
8449 4 : io_concurrency: IoConcurrency,
8450 4 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
8451 4 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8452 4 : let query = VersionedKeySpaceQuery::uniform(keyspace.clone(), lsn);
8453 4 : let res = tline
8454 4 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8455 4 : .await?;
8456 4 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
8457 4 : }
8458 :
8459 1001 : for blknum in 0..NUM_KEYS {
8460 1000 : lsn = Lsn(lsn.0 + 0x10);
8461 1000 : test_key.field6 = (blknum * STEP) as u32;
8462 1000 : let mut writer = tline.writer().await;
8463 1000 : writer
8464 1000 : .put(
8465 1000 : test_key,
8466 1000 : lsn,
8467 1000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8468 1000 : &ctx,
8469 1000 : )
8470 1000 : .await?;
8471 1000 : writer.finish_write(lsn);
8472 1000 : drop(writer);
8473 : }
8474 :
8475 1 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
8476 :
8477 11 : for iter in 1..=10 {
8478 10010 : for _ in 0..NUM_KEYS {
8479 10000 : lsn = Lsn(lsn.0 + 0x10);
8480 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
8481 10000 : test_key.field6 = (blknum * STEP) as u32;
8482 10000 : let mut writer = tline.writer().await;
8483 10000 : writer
8484 10000 : .put(
8485 10000 : test_key,
8486 10000 : lsn,
8487 10000 : &Value::Image(test_img(&format!("{blknum} at {lsn}"))),
8488 10000 : &ctx,
8489 10000 : )
8490 10000 : .await?;
8491 10000 : writer.finish_write(lsn);
8492 10000 : drop(writer);
8493 1 : }
8494 1 :
8495 10 : tline.freeze_and_flush().await?;
8496 1 : // Force layers to L1
8497 10 : tline
8498 10 : .compact(
8499 10 : &cancel,
8500 10 : {
8501 10 : let mut flags = EnumSet::new();
8502 10 : flags.insert(CompactFlags::ForceL0Compaction);
8503 10 : flags
8504 10 : },
8505 10 : &ctx,
8506 10 : )
8507 10 : .await?;
8508 1 :
8509 10 : if iter % 5 == 0 {
8510 2 : let scan_lsn = Lsn(lsn.0 + 1);
8511 2 : info!("scanning at {}", scan_lsn);
8512 2 : let (_, before_delta_file_accessed) =
8513 2 : scan_with_statistics(&tline, &keyspace, scan_lsn, &ctx, io_concurrency.clone())
8514 2 : .await?;
8515 2 : tline
8516 2 : .compact(
8517 2 : &cancel,
8518 2 : {
8519 2 : let mut flags = EnumSet::new();
8520 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
8521 2 : flags.insert(CompactFlags::ForceRepartition);
8522 2 : flags.insert(CompactFlags::ForceL0Compaction);
8523 2 : flags
8524 2 : },
8525 2 : &ctx,
8526 2 : )
8527 2 : .await?;
8528 2 : let (_, after_delta_file_accessed) =
8529 2 : scan_with_statistics(&tline, &keyspace, scan_lsn, &ctx, io_concurrency.clone())
8530 2 : .await?;
8531 2 : assert!(
8532 2 : after_delta_file_accessed < before_delta_file_accessed,
8533 1 : "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}"
8534 1 : );
8535 1 : // Given that we already produced an image layer, there should be no delta layer needed for the scan, but still setting a low threshold there for unforeseen circumstances.
8536 2 : assert!(
8537 2 : after_delta_file_accessed <= 2,
8538 1 : "after_delta_file_accessed={after_delta_file_accessed}"
8539 1 : );
8540 8 : }
8541 1 : }
8542 1 :
8543 1 : Ok(())
8544 1 : }
8545 :
8546 : #[tokio::test]
8547 1 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
8548 1 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
8549 1 : let (tenant, ctx) = harness.load().await;
8550 :
8551 1 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
8552 1 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
8553 1 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
8554 :
8555 1 : let tline = tenant
8556 1 : .create_test_timeline_with_layers(
8557 1 : TIMELINE_ID,
8558 1 : Lsn(0x10),
8559 1 : DEFAULT_PG_VERSION,
8560 1 : &ctx,
8561 1 : Vec::new(), // in-memory layers
8562 1 : Vec::new(), // delta layers
8563 1 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
8564 1 : Lsn(0x20), // it's fine to not advance LSN to 0x30 while using 0x30 to get below because `get_vectored_impl` does not wait for LSN
8565 1 : )
8566 1 : .await?;
8567 1 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
8568 :
8569 1 : let child = tenant
8570 1 : .branch_timeline_test_with_layers(
8571 1 : &tline,
8572 1 : NEW_TIMELINE_ID,
8573 1 : Some(Lsn(0x20)),
8574 1 : &ctx,
8575 1 : Vec::new(), // delta layers
8576 1 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
8577 1 : Lsn(0x30),
8578 1 : )
8579 1 : .await
8580 1 : .unwrap();
8581 :
8582 1 : let lsn = Lsn(0x30);
8583 :
8584 : // test vectored get on parent timeline
8585 1 : assert_eq!(
8586 1 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8587 1 : Some(test_img("data key 1"))
8588 : );
8589 1 : assert!(
8590 1 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
8591 1 : .await
8592 1 : .unwrap_err()
8593 1 : .is_missing_key_error()
8594 : );
8595 1 : assert!(
8596 1 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
8597 1 : .await
8598 1 : .unwrap_err()
8599 1 : .is_missing_key_error()
8600 : );
8601 :
8602 : // test vectored get on child timeline
8603 1 : assert_eq!(
8604 1 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8605 1 : Some(test_img("data key 1"))
8606 : );
8607 1 : assert_eq!(
8608 1 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8609 1 : Some(test_img("data key 2"))
8610 : );
8611 1 : assert!(
8612 1 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
8613 1 : .await
8614 1 : .unwrap_err()
8615 1 : .is_missing_key_error()
8616 : );
8617 :
8618 2 : Ok(())
8619 1 : }
8620 :
8621 : #[tokio::test]
8622 1 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
8623 1 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
8624 1 : let (tenant, ctx) = harness.load().await;
8625 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8626 :
8627 1 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
8628 1 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
8629 1 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
8630 1 : let base_key_overwrite = Key::from_hex("620000000033333333444444445500000003").unwrap();
8631 :
8632 1 : let base_inherited_key = Key::from_hex("610000000033333333444444445500000000").unwrap();
8633 1 : let base_inherited_key_child =
8634 1 : Key::from_hex("610000000033333333444444445500000001").unwrap();
8635 1 : let base_inherited_key_nonexist =
8636 1 : Key::from_hex("610000000033333333444444445500000002").unwrap();
8637 1 : let base_inherited_key_overwrite =
8638 1 : Key::from_hex("610000000033333333444444445500000003").unwrap();
8639 :
8640 1 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
8641 1 : assert_eq!(base_inherited_key.field1, RELATION_SIZE_PREFIX);
8642 :
8643 1 : let tline = tenant
8644 1 : .create_test_timeline_with_layers(
8645 1 : TIMELINE_ID,
8646 1 : Lsn(0x10),
8647 1 : DEFAULT_PG_VERSION,
8648 1 : &ctx,
8649 1 : Vec::new(), // in-memory layers
8650 1 : Vec::new(), // delta layers
8651 1 : vec![(
8652 1 : Lsn(0x20),
8653 1 : vec![
8654 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8655 1 : (
8656 1 : base_inherited_key_overwrite,
8657 1 : test_img("metadata key overwrite 1a"),
8658 1 : ),
8659 1 : (base_key, test_img("metadata key 1")),
8660 1 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8661 1 : ],
8662 1 : )], // image layers
8663 1 : Lsn(0x20), // it's fine to not advance LSN to 0x30 while using 0x30 to get below because `get_vectored_impl` does not wait for LSN
8664 1 : )
8665 1 : .await?;
8666 :
8667 1 : let child = tenant
8668 1 : .branch_timeline_test_with_layers(
8669 1 : &tline,
8670 1 : NEW_TIMELINE_ID,
8671 1 : Some(Lsn(0x20)),
8672 1 : &ctx,
8673 1 : Vec::new(), // delta layers
8674 1 : vec![(
8675 1 : Lsn(0x30),
8676 1 : vec![
8677 1 : (
8678 1 : base_inherited_key_child,
8679 1 : test_img("metadata inherited key 2"),
8680 1 : ),
8681 1 : (
8682 1 : base_inherited_key_overwrite,
8683 1 : test_img("metadata key overwrite 2a"),
8684 1 : ),
8685 1 : (base_key_child, test_img("metadata key 2")),
8686 1 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8687 1 : ],
8688 1 : )], // image layers
8689 1 : Lsn(0x30),
8690 1 : )
8691 1 : .await
8692 1 : .unwrap();
8693 :
8694 1 : let lsn = Lsn(0x30);
8695 :
8696 : // test vectored get on parent timeline
8697 1 : assert_eq!(
8698 1 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
8699 1 : Some(test_img("metadata key 1"))
8700 : );
8701 1 : assert_eq!(
8702 1 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
8703 : None
8704 : );
8705 1 : assert_eq!(
8706 1 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
8707 : None
8708 : );
8709 1 : assert_eq!(
8710 1 : get_vectored_impl_wrapper(&tline, base_key_overwrite, lsn, &ctx).await?,
8711 1 : Some(test_img("metadata key overwrite 1b"))
8712 : );
8713 1 : assert_eq!(
8714 1 : get_vectored_impl_wrapper(&tline, base_inherited_key, lsn, &ctx).await?,
8715 1 : Some(test_img("metadata inherited key 1"))
8716 : );
8717 1 : assert_eq!(
8718 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_child, lsn, &ctx).await?,
8719 : None
8720 : );
8721 1 : assert_eq!(
8722 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_nonexist, lsn, &ctx).await?,
8723 : None
8724 : );
8725 1 : assert_eq!(
8726 1 : get_vectored_impl_wrapper(&tline, base_inherited_key_overwrite, lsn, &ctx).await?,
8727 1 : Some(test_img("metadata key overwrite 1a"))
8728 : );
8729 :
8730 : // test vectored get on child timeline
8731 1 : assert_eq!(
8732 1 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
8733 : None
8734 : );
8735 1 : assert_eq!(
8736 1 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
8737 1 : Some(test_img("metadata key 2"))
8738 : );
8739 1 : assert_eq!(
8740 1 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
8741 : None
8742 : );
8743 1 : assert_eq!(
8744 1 : get_vectored_impl_wrapper(&child, base_inherited_key, lsn, &ctx).await?,
8745 1 : Some(test_img("metadata inherited key 1"))
8746 : );
8747 1 : assert_eq!(
8748 1 : get_vectored_impl_wrapper(&child, base_inherited_key_child, lsn, &ctx).await?,
8749 1 : Some(test_img("metadata inherited key 2"))
8750 : );
8751 1 : assert_eq!(
8752 1 : get_vectored_impl_wrapper(&child, base_inherited_key_nonexist, lsn, &ctx).await?,
8753 : None
8754 : );
8755 1 : assert_eq!(
8756 1 : get_vectored_impl_wrapper(&child, base_key_overwrite, lsn, &ctx).await?,
8757 1 : Some(test_img("metadata key overwrite 2b"))
8758 : );
8759 1 : assert_eq!(
8760 1 : get_vectored_impl_wrapper(&child, base_inherited_key_overwrite, lsn, &ctx).await?,
8761 1 : Some(test_img("metadata key overwrite 2a"))
8762 : );
8763 :
8764 : // test vectored scan on parent timeline
8765 1 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8766 1 : let query =
8767 1 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8768 1 : let res = tline
8769 1 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8770 1 : .await?;
8771 :
8772 1 : assert_eq!(
8773 1 : res.into_iter()
8774 4 : .map(|(k, v)| (k, v.unwrap()))
8775 1 : .collect::<Vec<_>>(),
8776 1 : vec![
8777 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8778 1 : (
8779 1 : base_inherited_key_overwrite,
8780 1 : test_img("metadata key overwrite 1a")
8781 1 : ),
8782 1 : (base_key, test_img("metadata key 1")),
8783 1 : (base_key_overwrite, test_img("metadata key overwrite 1b")),
8784 : ]
8785 : );
8786 :
8787 : // test vectored scan on child timeline
8788 1 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
8789 1 : let query =
8790 1 : VersionedKeySpaceQuery::uniform(KeySpace::single(Key::metadata_key_range()), lsn);
8791 1 : let res = child
8792 1 : .get_vectored_impl(query, &mut reconstruct_state, &ctx)
8793 1 : .await?;
8794 :
8795 1 : assert_eq!(
8796 1 : res.into_iter()
8797 5 : .map(|(k, v)| (k, v.unwrap()))
8798 1 : .collect::<Vec<_>>(),
8799 1 : vec![
8800 1 : (base_inherited_key, test_img("metadata inherited key 1")),
8801 1 : (
8802 1 : base_inherited_key_child,
8803 1 : test_img("metadata inherited key 2")
8804 1 : ),
8805 1 : (
8806 1 : base_inherited_key_overwrite,
8807 1 : test_img("metadata key overwrite 2a")
8808 1 : ),
8809 1 : (base_key_child, test_img("metadata key 2")),
8810 1 : (base_key_overwrite, test_img("metadata key overwrite 2b")),
8811 : ]
8812 : );
8813 :
8814 2 : Ok(())
8815 1 : }
8816 :
8817 28 : async fn get_vectored_impl_wrapper(
8818 28 : tline: &Arc<Timeline>,
8819 28 : key: Key,
8820 28 : lsn: Lsn,
8821 28 : ctx: &RequestContext,
8822 28 : ) -> Result<Option<Bytes>, GetVectoredError> {
8823 28 : let io_concurrency = IoConcurrency::spawn_from_conf(
8824 28 : tline.conf.get_vectored_concurrent_io,
8825 28 : tline.gate.enter().unwrap(),
8826 : );
8827 28 : let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
8828 28 : let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
8829 28 : let mut res = tline
8830 28 : .get_vectored_impl(query, &mut reconstruct_state, ctx)
8831 28 : .await?;
8832 25 : Ok(res.pop_last().map(|(k, v)| {
8833 16 : assert_eq!(k, key);
8834 16 : v.unwrap()
8835 16 : }))
8836 28 : }
8837 :
8838 : #[tokio::test]
8839 1 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
8840 1 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
8841 1 : let (tenant, ctx) = harness.load().await;
8842 1 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8843 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8844 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8845 1 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8846 :
8847 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
8848 : // Lsn 0x30 key0, key3, no key1+key2
8849 : // Lsn 0x20 key1+key2 tomestones
8850 : // Lsn 0x10 key1 in image, key2 in delta
8851 1 : let tline = tenant
8852 1 : .create_test_timeline_with_layers(
8853 1 : TIMELINE_ID,
8854 1 : Lsn(0x10),
8855 1 : DEFAULT_PG_VERSION,
8856 1 : &ctx,
8857 1 : Vec::new(), // in-memory layers
8858 1 : // delta layers
8859 1 : vec![
8860 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8861 1 : Lsn(0x10)..Lsn(0x20),
8862 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8863 1 : ),
8864 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8865 1 : Lsn(0x20)..Lsn(0x30),
8866 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8867 1 : ),
8868 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8869 1 : Lsn(0x20)..Lsn(0x30),
8870 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8871 1 : ),
8872 1 : ],
8873 1 : // image layers
8874 1 : vec![
8875 1 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
8876 1 : (
8877 1 : Lsn(0x30),
8878 1 : vec![
8879 1 : (key0, test_img("metadata key 0")),
8880 1 : (key3, test_img("metadata key 3")),
8881 1 : ],
8882 1 : ),
8883 1 : ],
8884 1 : Lsn(0x30),
8885 1 : )
8886 1 : .await?;
8887 :
8888 1 : let lsn = Lsn(0x30);
8889 1 : let old_lsn = Lsn(0x20);
8890 :
8891 1 : assert_eq!(
8892 1 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
8893 1 : Some(test_img("metadata key 0"))
8894 : );
8895 1 : assert_eq!(
8896 1 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
8897 : None,
8898 : );
8899 1 : assert_eq!(
8900 1 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
8901 : None,
8902 : );
8903 1 : assert_eq!(
8904 1 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
8905 1 : Some(Bytes::new()),
8906 : );
8907 1 : assert_eq!(
8908 1 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
8909 1 : Some(Bytes::new()),
8910 : );
8911 1 : assert_eq!(
8912 1 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
8913 1 : Some(test_img("metadata key 3"))
8914 : );
8915 :
8916 2 : Ok(())
8917 1 : }
8918 :
8919 : #[tokio::test]
8920 1 : async fn test_metadata_tombstone_image_creation() {
8921 1 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
8922 1 : .await
8923 1 : .unwrap();
8924 1 : let (tenant, ctx) = harness.load().await;
8925 1 : let io_concurrency = IoConcurrency::spawn_for_test();
8926 :
8927 1 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
8928 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
8929 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
8930 1 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
8931 :
8932 1 : let tline = tenant
8933 1 : .create_test_timeline_with_layers(
8934 1 : TIMELINE_ID,
8935 1 : Lsn(0x10),
8936 1 : DEFAULT_PG_VERSION,
8937 1 : &ctx,
8938 1 : Vec::new(), // in-memory layers
8939 1 : // delta layers
8940 1 : vec![
8941 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8942 1 : Lsn(0x10)..Lsn(0x20),
8943 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
8944 1 : ),
8945 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8946 1 : Lsn(0x20)..Lsn(0x30),
8947 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
8948 1 : ),
8949 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8950 1 : Lsn(0x20)..Lsn(0x30),
8951 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
8952 1 : ),
8953 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
8954 1 : Lsn(0x30)..Lsn(0x40),
8955 1 : vec![
8956 1 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
8957 1 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
8958 1 : ],
8959 1 : ),
8960 1 : ],
8961 1 : // image layers
8962 1 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
8963 1 : Lsn(0x40),
8964 1 : )
8965 1 : .await
8966 1 : .unwrap();
8967 :
8968 1 : let cancel = CancellationToken::new();
8969 :
8970 : // Image layer creation happens on the disk_consistent_lsn so we need to force set it now.
8971 1 : tline.force_set_disk_consistent_lsn(Lsn(0x40));
8972 1 : tline
8973 1 : .compact(
8974 1 : &cancel,
8975 1 : {
8976 1 : let mut flags = EnumSet::new();
8977 1 : flags.insert(CompactFlags::ForceImageLayerCreation);
8978 1 : flags.insert(CompactFlags::ForceRepartition);
8979 1 : flags
8980 1 : },
8981 1 : &ctx,
8982 1 : )
8983 1 : .await
8984 1 : .unwrap();
8985 : // Image layers are created at repartition LSN
8986 1 : let images = tline
8987 1 : .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
8988 1 : .await
8989 1 : .unwrap()
8990 1 : .into_iter()
8991 9 : .filter(|(k, _)| k.is_metadata_key())
8992 1 : .collect::<Vec<_>>();
8993 1 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
8994 1 : }
8995 :
8996 : #[tokio::test]
8997 1 : async fn test_metadata_tombstone_empty_image_creation() {
8998 1 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
8999 1 : .await
9000 1 : .unwrap();
9001 1 : let (tenant, ctx) = harness.load().await;
9002 1 : let io_concurrency = IoConcurrency::spawn_for_test();
9003 :
9004 1 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
9005 1 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
9006 :
9007 1 : let tline = tenant
9008 1 : .create_test_timeline_with_layers(
9009 1 : TIMELINE_ID,
9010 1 : Lsn(0x10),
9011 1 : DEFAULT_PG_VERSION,
9012 1 : &ctx,
9013 1 : Vec::new(), // in-memory layers
9014 1 : // delta layers
9015 1 : vec![
9016 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9017 1 : Lsn(0x10)..Lsn(0x20),
9018 1 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
9019 1 : ),
9020 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9021 1 : Lsn(0x20)..Lsn(0x30),
9022 1 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
9023 1 : ),
9024 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9025 1 : Lsn(0x20)..Lsn(0x30),
9026 1 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
9027 1 : ),
9028 1 : ],
9029 1 : // image layers
9030 1 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
9031 1 : Lsn(0x30),
9032 1 : )
9033 1 : .await
9034 1 : .unwrap();
9035 :
9036 1 : let cancel = CancellationToken::new();
9037 :
9038 1 : tline
9039 1 : .compact(
9040 1 : &cancel,
9041 1 : {
9042 1 : let mut flags = EnumSet::new();
9043 1 : flags.insert(CompactFlags::ForceImageLayerCreation);
9044 1 : flags.insert(CompactFlags::ForceRepartition);
9045 1 : flags
9046 1 : },
9047 1 : &ctx,
9048 1 : )
9049 1 : .await
9050 1 : .unwrap();
9051 :
9052 : // Image layers are created at last_record_lsn
9053 1 : let images = tline
9054 1 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
9055 1 : .await
9056 1 : .unwrap()
9057 1 : .into_iter()
9058 7 : .filter(|(k, _)| k.is_metadata_key())
9059 1 : .collect::<Vec<_>>();
9060 1 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
9061 1 : }
9062 :
9063 : #[tokio::test]
9064 1 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
9065 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
9066 1 : let (tenant, ctx) = harness.load().await;
9067 1 : let io_concurrency = IoConcurrency::spawn_for_test();
9068 :
9069 51 : fn get_key(id: u32) -> Key {
9070 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9071 51 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9072 51 : key.field6 = id;
9073 51 : key
9074 51 : }
9075 :
9076 : // We create
9077 : // - one bottom-most image layer,
9078 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
9079 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
9080 : // - a delta layer D3 above the horizon.
9081 : //
9082 : // | D3 |
9083 : // | D1 |
9084 : // -| |-- gc horizon -----------------
9085 : // | | | D2 |
9086 : // --------- img layer ------------------
9087 : //
9088 : // What we should expact from this compaction is:
9089 : // | D3 |
9090 : // | Part of D1 |
9091 : // --------- img layer with D1+D2 at GC horizon------------------
9092 :
9093 : // img layer at 0x10
9094 1 : let img_layer = (0..10)
9095 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9096 1 : .collect_vec();
9097 :
9098 1 : let delta1 = vec![
9099 1 : (
9100 1 : get_key(1),
9101 1 : Lsn(0x20),
9102 1 : Value::Image(Bytes::from("value 1@0x20")),
9103 1 : ),
9104 1 : (
9105 1 : get_key(2),
9106 1 : Lsn(0x30),
9107 1 : Value::Image(Bytes::from("value 2@0x30")),
9108 1 : ),
9109 1 : (
9110 1 : get_key(3),
9111 1 : Lsn(0x40),
9112 1 : Value::Image(Bytes::from("value 3@0x40")),
9113 1 : ),
9114 : ];
9115 1 : let delta2 = vec![
9116 1 : (
9117 1 : get_key(5),
9118 1 : Lsn(0x20),
9119 1 : Value::Image(Bytes::from("value 5@0x20")),
9120 1 : ),
9121 1 : (
9122 1 : get_key(6),
9123 1 : Lsn(0x20),
9124 1 : Value::Image(Bytes::from("value 6@0x20")),
9125 1 : ),
9126 : ];
9127 1 : let delta3 = vec![
9128 1 : (
9129 1 : get_key(8),
9130 1 : Lsn(0x48),
9131 1 : Value::Image(Bytes::from("value 8@0x48")),
9132 1 : ),
9133 1 : (
9134 1 : get_key(9),
9135 1 : Lsn(0x48),
9136 1 : Value::Image(Bytes::from("value 9@0x48")),
9137 1 : ),
9138 : ];
9139 :
9140 1 : let tline = tenant
9141 1 : .create_test_timeline_with_layers(
9142 1 : TIMELINE_ID,
9143 1 : Lsn(0x10),
9144 1 : DEFAULT_PG_VERSION,
9145 1 : &ctx,
9146 1 : Vec::new(), // in-memory layers
9147 1 : vec![
9148 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
9149 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
9150 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
9151 1 : ], // delta layers
9152 1 : vec![(Lsn(0x10), img_layer)], // image layers
9153 1 : Lsn(0x50),
9154 1 : )
9155 1 : .await?;
9156 : {
9157 1 : tline
9158 1 : .applied_gc_cutoff_lsn
9159 1 : .lock_for_write()
9160 1 : .store_and_unlock(Lsn(0x30))
9161 1 : .wait()
9162 1 : .await;
9163 : // Update GC info
9164 1 : let mut guard = tline.gc_info.write().unwrap();
9165 1 : guard.cutoffs.time = Some(Lsn(0x30));
9166 1 : guard.cutoffs.space = Lsn(0x30);
9167 : }
9168 :
9169 1 : let expected_result = [
9170 1 : Bytes::from_static(b"value 0@0x10"),
9171 1 : Bytes::from_static(b"value 1@0x20"),
9172 1 : Bytes::from_static(b"value 2@0x30"),
9173 1 : Bytes::from_static(b"value 3@0x40"),
9174 1 : Bytes::from_static(b"value 4@0x10"),
9175 1 : Bytes::from_static(b"value 5@0x20"),
9176 1 : Bytes::from_static(b"value 6@0x20"),
9177 1 : Bytes::from_static(b"value 7@0x10"),
9178 1 : Bytes::from_static(b"value 8@0x48"),
9179 1 : Bytes::from_static(b"value 9@0x48"),
9180 1 : ];
9181 :
9182 10 : for (idx, expected) in expected_result.iter().enumerate() {
9183 10 : assert_eq!(
9184 10 : tline
9185 10 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9186 10 : .await
9187 10 : .unwrap(),
9188 : expected
9189 : );
9190 : }
9191 :
9192 1 : let cancel = CancellationToken::new();
9193 1 : tline
9194 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9195 1 : .await
9196 1 : .unwrap();
9197 :
9198 10 : for (idx, expected) in expected_result.iter().enumerate() {
9199 10 : assert_eq!(
9200 10 : tline
9201 10 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9202 10 : .await
9203 10 : .unwrap(),
9204 : expected
9205 : );
9206 : }
9207 :
9208 : // Check if the image layer at the GC horizon contains exactly what we want
9209 1 : let image_at_gc_horizon = tline
9210 1 : .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
9211 1 : .await
9212 1 : .unwrap()
9213 1 : .into_iter()
9214 17 : .filter(|(k, _)| k.is_metadata_key())
9215 1 : .collect::<Vec<_>>();
9216 :
9217 1 : assert_eq!(image_at_gc_horizon.len(), 10);
9218 1 : let expected_result = [
9219 1 : Bytes::from_static(b"value 0@0x10"),
9220 1 : Bytes::from_static(b"value 1@0x20"),
9221 1 : Bytes::from_static(b"value 2@0x30"),
9222 1 : Bytes::from_static(b"value 3@0x10"),
9223 1 : Bytes::from_static(b"value 4@0x10"),
9224 1 : Bytes::from_static(b"value 5@0x20"),
9225 1 : Bytes::from_static(b"value 6@0x20"),
9226 1 : Bytes::from_static(b"value 7@0x10"),
9227 1 : Bytes::from_static(b"value 8@0x10"),
9228 1 : Bytes::from_static(b"value 9@0x10"),
9229 1 : ];
9230 11 : for idx in 0..10 {
9231 10 : assert_eq!(
9232 10 : image_at_gc_horizon[idx],
9233 10 : (get_key(idx as u32), expected_result[idx].clone())
9234 : );
9235 : }
9236 :
9237 : // Check if old layers are removed / new layers have the expected LSN
9238 1 : let all_layers = inspect_and_sort(&tline, None).await;
9239 1 : assert_eq!(
9240 : all_layers,
9241 1 : vec![
9242 : // Image layer at GC horizon
9243 1 : PersistentLayerKey {
9244 1 : key_range: Key::MIN..Key::MAX,
9245 1 : lsn_range: Lsn(0x30)..Lsn(0x31),
9246 1 : is_delta: false
9247 1 : },
9248 : // The delta layer below the horizon
9249 1 : PersistentLayerKey {
9250 1 : key_range: get_key(3)..get_key(4),
9251 1 : lsn_range: Lsn(0x30)..Lsn(0x48),
9252 1 : is_delta: true
9253 1 : },
9254 : // The delta3 layer that should not be picked for the compaction
9255 1 : PersistentLayerKey {
9256 1 : key_range: get_key(8)..get_key(10),
9257 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
9258 1 : is_delta: true
9259 1 : }
9260 : ]
9261 : );
9262 :
9263 : // increase GC horizon and compact again
9264 : {
9265 1 : tline
9266 1 : .applied_gc_cutoff_lsn
9267 1 : .lock_for_write()
9268 1 : .store_and_unlock(Lsn(0x40))
9269 1 : .wait()
9270 1 : .await;
9271 : // Update GC info
9272 1 : let mut guard = tline.gc_info.write().unwrap();
9273 1 : guard.cutoffs.time = Some(Lsn(0x40));
9274 1 : guard.cutoffs.space = Lsn(0x40);
9275 : }
9276 1 : tline
9277 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9278 1 : .await
9279 1 : .unwrap();
9280 :
9281 2 : Ok(())
9282 1 : }
9283 :
9284 : #[cfg(feature = "testing")]
9285 : #[tokio::test]
9286 1 : async fn test_neon_test_record() -> anyhow::Result<()> {
9287 1 : let harness = TenantHarness::create("test_neon_test_record").await?;
9288 1 : let (tenant, ctx) = harness.load().await;
9289 :
9290 17 : fn get_key(id: u32) -> Key {
9291 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9292 17 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9293 17 : key.field6 = id;
9294 17 : key
9295 17 : }
9296 :
9297 1 : let delta1 = vec![
9298 1 : (
9299 1 : get_key(1),
9300 1 : Lsn(0x20),
9301 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9302 1 : ),
9303 1 : (
9304 1 : get_key(1),
9305 1 : Lsn(0x30),
9306 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9307 1 : ),
9308 1 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
9309 1 : (
9310 1 : get_key(2),
9311 1 : Lsn(0x20),
9312 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
9313 1 : ),
9314 1 : (
9315 1 : get_key(2),
9316 1 : Lsn(0x30),
9317 1 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
9318 1 : ),
9319 1 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
9320 1 : (
9321 1 : get_key(3),
9322 1 : Lsn(0x20),
9323 1 : Value::WalRecord(NeonWalRecord::wal_clear("c")),
9324 1 : ),
9325 1 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
9326 1 : (
9327 1 : get_key(4),
9328 1 : Lsn(0x20),
9329 1 : Value::WalRecord(NeonWalRecord::wal_init("i")),
9330 1 : ),
9331 1 : (
9332 1 : get_key(4),
9333 1 : Lsn(0x30),
9334 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "i")),
9335 1 : ),
9336 1 : (
9337 1 : get_key(5),
9338 1 : Lsn(0x20),
9339 1 : Value::WalRecord(NeonWalRecord::wal_init("1")),
9340 1 : ),
9341 1 : (
9342 1 : get_key(5),
9343 1 : Lsn(0x30),
9344 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "2")),
9345 1 : ),
9346 : ];
9347 1 : let image1 = vec![(get_key(1), "0x10".into())];
9348 :
9349 1 : let tline = tenant
9350 1 : .create_test_timeline_with_layers(
9351 1 : TIMELINE_ID,
9352 1 : Lsn(0x10),
9353 1 : DEFAULT_PG_VERSION,
9354 1 : &ctx,
9355 1 : Vec::new(), // in-memory layers
9356 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
9357 1 : Lsn(0x10)..Lsn(0x40),
9358 1 : delta1,
9359 1 : )], // delta layers
9360 1 : vec![(Lsn(0x10), image1)], // image layers
9361 1 : Lsn(0x50),
9362 1 : )
9363 1 : .await?;
9364 :
9365 1 : assert_eq!(
9366 1 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
9367 1 : Bytes::from_static(b"0x10,0x20,0x30")
9368 : );
9369 1 : assert_eq!(
9370 1 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
9371 1 : Bytes::from_static(b"0x10,0x20,0x30")
9372 : );
9373 :
9374 : // Need to remove the limit of "Neon WAL redo requires base image".
9375 :
9376 1 : assert_eq!(
9377 1 : tline.get(get_key(3), Lsn(0x50), &ctx).await?,
9378 1 : Bytes::from_static(b"c")
9379 : );
9380 1 : assert_eq!(
9381 1 : tline.get(get_key(4), Lsn(0x50), &ctx).await?,
9382 1 : Bytes::from_static(b"ij")
9383 : );
9384 :
9385 : // Manual testing required: currently, read errors will panic the process in debug mode. So we
9386 : // cannot enable this assertion in the unit test.
9387 : // assert!(tline.get(get_key(5), Lsn(0x50), &ctx).await.is_err());
9388 :
9389 2 : Ok(())
9390 1 : }
9391 :
9392 : #[tokio::test]
9393 1 : async fn test_lsn_lease() -> anyhow::Result<()> {
9394 1 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
9395 1 : .await
9396 1 : .unwrap()
9397 1 : .load()
9398 1 : .await;
9399 : // set a non-zero lease length to test the feature
9400 1 : tenant
9401 1 : .update_tenant_config(|mut conf| {
9402 1 : conf.lsn_lease_length = Some(LsnLease::DEFAULT_LENGTH);
9403 1 : Ok(conf)
9404 1 : })
9405 1 : .unwrap();
9406 :
9407 1 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9408 :
9409 1 : let end_lsn = Lsn(0x100);
9410 1 : let image_layers = (0x20..=0x90)
9411 1 : .step_by(0x10)
9412 8 : .map(|n| (Lsn(n), vec![(key, test_img(&format!("data key at {n:x}")))]))
9413 1 : .collect();
9414 :
9415 1 : let timeline = tenant
9416 1 : .create_test_timeline_with_layers(
9417 1 : TIMELINE_ID,
9418 1 : Lsn(0x10),
9419 1 : DEFAULT_PG_VERSION,
9420 1 : &ctx,
9421 1 : Vec::new(), // in-memory layers
9422 1 : Vec::new(),
9423 1 : image_layers,
9424 1 : end_lsn,
9425 1 : )
9426 1 : .await?;
9427 :
9428 1 : let leased_lsns = [0x30, 0x50, 0x70];
9429 1 : let mut leases = Vec::new();
9430 3 : leased_lsns.iter().for_each(|n| {
9431 3 : leases.push(
9432 3 : timeline
9433 3 : .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
9434 3 : .expect("lease request should succeed"),
9435 : );
9436 3 : });
9437 :
9438 1 : let updated_lease_0 = timeline
9439 1 : .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
9440 1 : .expect("lease renewal should succeed");
9441 1 : assert_eq!(
9442 1 : updated_lease_0.valid_until, leases[0].valid_until,
9443 0 : " Renewing with shorter lease should not change the lease."
9444 : );
9445 :
9446 1 : let updated_lease_1 = timeline
9447 1 : .renew_lsn_lease(
9448 1 : Lsn(leased_lsns[1]),
9449 1 : timeline.get_lsn_lease_length() * 2,
9450 1 : &ctx,
9451 1 : )
9452 1 : .expect("lease renewal should succeed");
9453 1 : assert!(
9454 1 : updated_lease_1.valid_until > leases[1].valid_until,
9455 0 : "Renewing with a long lease should renew lease with later expiration time."
9456 : );
9457 :
9458 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
9459 1 : info!(
9460 0 : "applied_gc_cutoff_lsn: {}",
9461 0 : *timeline.get_applied_gc_cutoff_lsn()
9462 : );
9463 1 : timeline.force_set_disk_consistent_lsn(end_lsn);
9464 :
9465 1 : let res = tenant
9466 1 : .gc_iteration(
9467 1 : Some(TIMELINE_ID),
9468 1 : 0,
9469 1 : Duration::ZERO,
9470 1 : &CancellationToken::new(),
9471 1 : &ctx,
9472 1 : )
9473 1 : .await
9474 1 : .unwrap();
9475 :
9476 : // Keeping everything <= Lsn(0x80) b/c leases:
9477 : // 0/10: initdb layer
9478 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
9479 1 : assert_eq!(res.layers_needed_by_leases, 7);
9480 : // Keeping 0/90 b/c it is the latest layer.
9481 1 : assert_eq!(res.layers_not_updated, 1);
9482 : // Removed 0/80.
9483 1 : assert_eq!(res.layers_removed, 1);
9484 :
9485 : // Make lease on a already GC-ed LSN.
9486 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
9487 1 : assert!(Lsn(0x80) < *timeline.get_applied_gc_cutoff_lsn());
9488 1 : timeline
9489 1 : .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
9490 1 : .expect_err("lease request on GC-ed LSN should fail");
9491 :
9492 : // Should still be able to renew a currently valid lease
9493 : // Assumption: original lease to is still valid for 0/50.
9494 : // (use `Timeline::init_lsn_lease` for testing so it always does validation)
9495 1 : timeline
9496 1 : .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
9497 1 : .expect("lease renewal with validation should succeed");
9498 :
9499 2 : Ok(())
9500 1 : }
9501 :
9502 : #[tokio::test]
9503 1 : async fn test_failed_flush_should_not_update_disk_consistent_lsn() -> anyhow::Result<()> {
9504 : //
9505 : // Setup
9506 : //
9507 1 : let harness = TenantHarness::create_custom(
9508 1 : "test_failed_flush_should_not_upload_disk_consistent_lsn",
9509 1 : pageserver_api::models::TenantConfig::default(),
9510 1 : TenantId::generate(),
9511 1 : ShardIdentity::new(ShardNumber(0), ShardCount(4), ShardStripeSize(128)).unwrap(),
9512 1 : Generation::new(1),
9513 1 : )
9514 1 : .await?;
9515 1 : let (tenant, ctx) = harness.load().await;
9516 :
9517 1 : let timeline = tenant
9518 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9519 1 : .await?;
9520 1 : assert_eq!(timeline.get_shard_identity().count, ShardCount(4));
9521 1 : let mut writer = timeline.writer().await;
9522 1 : writer
9523 1 : .put(
9524 1 : *TEST_KEY,
9525 1 : Lsn(0x20),
9526 1 : &Value::Image(test_img("foo at 0x20")),
9527 1 : &ctx,
9528 1 : )
9529 1 : .await?;
9530 1 : writer.finish_write(Lsn(0x20));
9531 1 : drop(writer);
9532 1 : timeline.freeze_and_flush().await.unwrap();
9533 :
9534 1 : timeline.remote_client.wait_completion().await.unwrap();
9535 1 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
9536 1 : let remote_consistent_lsn = timeline.get_remote_consistent_lsn_projected();
9537 1 : assert_eq!(Some(disk_consistent_lsn), remote_consistent_lsn);
9538 :
9539 : //
9540 : // Test
9541 : //
9542 :
9543 1 : let mut writer = timeline.writer().await;
9544 1 : writer
9545 1 : .put(
9546 1 : *TEST_KEY,
9547 1 : Lsn(0x30),
9548 1 : &Value::Image(test_img("foo at 0x30")),
9549 1 : &ctx,
9550 1 : )
9551 1 : .await?;
9552 1 : writer.finish_write(Lsn(0x30));
9553 1 : drop(writer);
9554 :
9555 1 : fail::cfg(
9556 : "flush-layer-before-update-remote-consistent-lsn",
9557 1 : "return()",
9558 : )
9559 1 : .unwrap();
9560 :
9561 1 : let flush_res = timeline.freeze_and_flush().await;
9562 : // if flush failed, the disk/remote consistent LSN should not be updated
9563 1 : assert!(flush_res.is_err());
9564 1 : assert_eq!(disk_consistent_lsn, timeline.get_disk_consistent_lsn());
9565 1 : assert_eq!(
9566 : remote_consistent_lsn,
9567 1 : timeline.get_remote_consistent_lsn_projected()
9568 : );
9569 :
9570 2 : Ok(())
9571 1 : }
9572 :
9573 : #[cfg(feature = "testing")]
9574 : #[tokio::test]
9575 1 : async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
9576 2 : test_simple_bottom_most_compaction_deltas_helper(
9577 2 : "test_simple_bottom_most_compaction_deltas_1",
9578 2 : false,
9579 2 : )
9580 2 : .await
9581 1 : }
9582 :
9583 : #[cfg(feature = "testing")]
9584 : #[tokio::test]
9585 1 : async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
9586 2 : test_simple_bottom_most_compaction_deltas_helper(
9587 2 : "test_simple_bottom_most_compaction_deltas_2",
9588 2 : true,
9589 2 : )
9590 2 : .await
9591 1 : }
9592 :
9593 : #[cfg(feature = "testing")]
9594 2 : async fn test_simple_bottom_most_compaction_deltas_helper(
9595 2 : test_name: &'static str,
9596 2 : use_delta_bottom_layer: bool,
9597 2 : ) -> anyhow::Result<()> {
9598 2 : let harness = TenantHarness::create(test_name).await?;
9599 2 : let (tenant, ctx) = harness.load().await;
9600 :
9601 138 : fn get_key(id: u32) -> Key {
9602 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
9603 138 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
9604 138 : key.field6 = id;
9605 138 : key
9606 138 : }
9607 :
9608 : // We create
9609 : // - one bottom-most image layer,
9610 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
9611 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
9612 : // - a delta layer D3 above the horizon.
9613 : //
9614 : // | D3 |
9615 : // | D1 |
9616 : // -| |-- gc horizon -----------------
9617 : // | | | D2 |
9618 : // --------- img layer ------------------
9619 : //
9620 : // What we should expact from this compaction is:
9621 : // | D3 |
9622 : // | Part of D1 |
9623 : // --------- img layer with D1+D2 at GC horizon------------------
9624 :
9625 : // img layer at 0x10
9626 2 : let img_layer = (0..10)
9627 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
9628 2 : .collect_vec();
9629 : // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
9630 2 : let delta4 = (0..10)
9631 20 : .map(|id| {
9632 20 : (
9633 20 : get_key(id),
9634 20 : Lsn(0x08),
9635 20 : Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
9636 20 : )
9637 20 : })
9638 2 : .collect_vec();
9639 :
9640 2 : let delta1 = vec![
9641 2 : (
9642 2 : get_key(1),
9643 2 : Lsn(0x20),
9644 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9645 2 : ),
9646 2 : (
9647 2 : get_key(2),
9648 2 : Lsn(0x30),
9649 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9650 2 : ),
9651 2 : (
9652 2 : get_key(3),
9653 2 : Lsn(0x28),
9654 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
9655 2 : ),
9656 2 : (
9657 2 : get_key(3),
9658 2 : Lsn(0x30),
9659 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
9660 2 : ),
9661 2 : (
9662 2 : get_key(3),
9663 2 : Lsn(0x40),
9664 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
9665 2 : ),
9666 : ];
9667 2 : let delta2 = vec![
9668 2 : (
9669 2 : get_key(5),
9670 2 : Lsn(0x20),
9671 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9672 2 : ),
9673 2 : (
9674 2 : get_key(6),
9675 2 : Lsn(0x20),
9676 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
9677 2 : ),
9678 : ];
9679 2 : let delta3 = vec![
9680 2 : (
9681 2 : get_key(8),
9682 2 : Lsn(0x48),
9683 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9684 2 : ),
9685 2 : (
9686 2 : get_key(9),
9687 2 : Lsn(0x48),
9688 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
9689 2 : ),
9690 : ];
9691 :
9692 2 : let tline = if use_delta_bottom_layer {
9693 1 : tenant
9694 1 : .create_test_timeline_with_layers(
9695 1 : TIMELINE_ID,
9696 1 : Lsn(0x08),
9697 1 : DEFAULT_PG_VERSION,
9698 1 : &ctx,
9699 1 : Vec::new(), // in-memory layers
9700 1 : vec![
9701 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9702 1 : Lsn(0x08)..Lsn(0x10),
9703 1 : delta4,
9704 1 : ),
9705 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9706 1 : Lsn(0x20)..Lsn(0x48),
9707 1 : delta1,
9708 1 : ),
9709 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9710 1 : Lsn(0x20)..Lsn(0x48),
9711 1 : delta2,
9712 1 : ),
9713 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9714 1 : Lsn(0x48)..Lsn(0x50),
9715 1 : delta3,
9716 1 : ),
9717 1 : ], // delta layers
9718 1 : vec![], // image layers
9719 1 : Lsn(0x50),
9720 1 : )
9721 1 : .await?
9722 : } else {
9723 1 : tenant
9724 1 : .create_test_timeline_with_layers(
9725 1 : TIMELINE_ID,
9726 1 : Lsn(0x10),
9727 1 : DEFAULT_PG_VERSION,
9728 1 : &ctx,
9729 1 : Vec::new(), // in-memory layers
9730 1 : vec![
9731 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9732 1 : Lsn(0x10)..Lsn(0x48),
9733 1 : delta1,
9734 1 : ),
9735 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9736 1 : Lsn(0x10)..Lsn(0x48),
9737 1 : delta2,
9738 1 : ),
9739 1 : DeltaLayerTestDesc::new_with_inferred_key_range(
9740 1 : Lsn(0x48)..Lsn(0x50),
9741 1 : delta3,
9742 1 : ),
9743 1 : ], // delta layers
9744 1 : vec![(Lsn(0x10), img_layer)], // image layers
9745 1 : Lsn(0x50),
9746 1 : )
9747 1 : .await?
9748 : };
9749 : {
9750 2 : tline
9751 2 : .applied_gc_cutoff_lsn
9752 2 : .lock_for_write()
9753 2 : .store_and_unlock(Lsn(0x30))
9754 2 : .wait()
9755 2 : .await;
9756 : // Update GC info
9757 2 : let mut guard = tline.gc_info.write().unwrap();
9758 2 : *guard = GcInfo {
9759 2 : retain_lsns: vec![],
9760 2 : cutoffs: GcCutoffs {
9761 2 : time: Some(Lsn(0x30)),
9762 2 : space: Lsn(0x30),
9763 2 : },
9764 2 : leases: Default::default(),
9765 2 : within_ancestor_pitr: false,
9766 2 : };
9767 : }
9768 :
9769 2 : let expected_result = [
9770 2 : Bytes::from_static(b"value 0@0x10"),
9771 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9772 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9773 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
9774 2 : Bytes::from_static(b"value 4@0x10"),
9775 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9776 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9777 2 : Bytes::from_static(b"value 7@0x10"),
9778 2 : Bytes::from_static(b"value 8@0x10@0x48"),
9779 2 : Bytes::from_static(b"value 9@0x10@0x48"),
9780 2 : ];
9781 :
9782 2 : let expected_result_at_gc_horizon = [
9783 2 : Bytes::from_static(b"value 0@0x10"),
9784 2 : Bytes::from_static(b"value 1@0x10@0x20"),
9785 2 : Bytes::from_static(b"value 2@0x10@0x30"),
9786 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
9787 2 : Bytes::from_static(b"value 4@0x10"),
9788 2 : Bytes::from_static(b"value 5@0x10@0x20"),
9789 2 : Bytes::from_static(b"value 6@0x10@0x20"),
9790 2 : Bytes::from_static(b"value 7@0x10"),
9791 2 : Bytes::from_static(b"value 8@0x10"),
9792 2 : Bytes::from_static(b"value 9@0x10"),
9793 2 : ];
9794 :
9795 22 : for idx in 0..10 {
9796 20 : assert_eq!(
9797 20 : tline
9798 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9799 20 : .await
9800 20 : .unwrap(),
9801 20 : &expected_result[idx]
9802 : );
9803 20 : assert_eq!(
9804 20 : tline
9805 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9806 20 : .await
9807 20 : .unwrap(),
9808 20 : &expected_result_at_gc_horizon[idx]
9809 : );
9810 : }
9811 :
9812 2 : let cancel = CancellationToken::new();
9813 2 : tline
9814 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9815 2 : .await
9816 2 : .unwrap();
9817 :
9818 22 : for idx in 0..10 {
9819 20 : assert_eq!(
9820 20 : tline
9821 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
9822 20 : .await
9823 20 : .unwrap(),
9824 20 : &expected_result[idx]
9825 : );
9826 20 : assert_eq!(
9827 20 : tline
9828 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
9829 20 : .await
9830 20 : .unwrap(),
9831 20 : &expected_result_at_gc_horizon[idx]
9832 : );
9833 : }
9834 :
9835 : // increase GC horizon and compact again
9836 : {
9837 2 : tline
9838 2 : .applied_gc_cutoff_lsn
9839 2 : .lock_for_write()
9840 2 : .store_and_unlock(Lsn(0x40))
9841 2 : .wait()
9842 2 : .await;
9843 : // Update GC info
9844 2 : let mut guard = tline.gc_info.write().unwrap();
9845 2 : guard.cutoffs.time = Some(Lsn(0x40));
9846 2 : guard.cutoffs.space = Lsn(0x40);
9847 : }
9848 2 : tline
9849 2 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
9850 2 : .await
9851 2 : .unwrap();
9852 :
9853 2 : Ok(())
9854 2 : }
9855 :
9856 : #[cfg(feature = "testing")]
9857 : #[tokio::test]
9858 1 : async fn test_generate_key_retention() -> anyhow::Result<()> {
9859 1 : let harness = TenantHarness::create("test_generate_key_retention").await?;
9860 1 : let (tenant, ctx) = harness.load().await;
9861 1 : let tline = tenant
9862 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
9863 1 : .await?;
9864 1 : tline.force_advance_lsn(Lsn(0x70));
9865 1 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
9866 1 : let history = vec![
9867 1 : (
9868 1 : key,
9869 1 : Lsn(0x10),
9870 1 : Value::WalRecord(NeonWalRecord::wal_init("0x10")),
9871 1 : ),
9872 1 : (
9873 1 : key,
9874 1 : Lsn(0x20),
9875 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
9876 1 : ),
9877 1 : (
9878 1 : key,
9879 1 : Lsn(0x30),
9880 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9881 1 : ),
9882 1 : (
9883 1 : key,
9884 1 : Lsn(0x40),
9885 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9886 1 : ),
9887 1 : (
9888 1 : key,
9889 1 : Lsn(0x50),
9890 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
9891 1 : ),
9892 1 : (
9893 1 : key,
9894 1 : Lsn(0x60),
9895 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9896 1 : ),
9897 1 : (
9898 1 : key,
9899 1 : Lsn(0x70),
9900 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9901 1 : ),
9902 1 : (
9903 1 : key,
9904 1 : Lsn(0x80),
9905 1 : Value::Image(Bytes::copy_from_slice(
9906 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9907 1 : )),
9908 1 : ),
9909 1 : (
9910 1 : key,
9911 1 : Lsn(0x90),
9912 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9913 1 : ),
9914 : ];
9915 1 : let res = tline
9916 1 : .generate_key_retention(
9917 1 : key,
9918 1 : &history,
9919 1 : Lsn(0x60),
9920 1 : &[Lsn(0x20), Lsn(0x40), Lsn(0x50)],
9921 1 : 3,
9922 1 : None,
9923 1 : true,
9924 1 : )
9925 1 : .await
9926 1 : .unwrap();
9927 1 : let expected_res = KeyHistoryRetention {
9928 1 : below_horizon: vec![
9929 1 : (
9930 1 : Lsn(0x20),
9931 1 : KeyLogAtLsn(vec![(
9932 1 : Lsn(0x20),
9933 1 : Value::Image(Bytes::from_static(b"0x10;0x20")),
9934 1 : )]),
9935 1 : ),
9936 1 : (
9937 1 : Lsn(0x40),
9938 1 : KeyLogAtLsn(vec![
9939 1 : (
9940 1 : Lsn(0x30),
9941 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
9942 1 : ),
9943 1 : (
9944 1 : Lsn(0x40),
9945 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
9946 1 : ),
9947 1 : ]),
9948 1 : ),
9949 1 : (
9950 1 : Lsn(0x50),
9951 1 : KeyLogAtLsn(vec![(
9952 1 : Lsn(0x50),
9953 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40;0x50")),
9954 1 : )]),
9955 1 : ),
9956 1 : (
9957 1 : Lsn(0x60),
9958 1 : KeyLogAtLsn(vec![(
9959 1 : Lsn(0x60),
9960 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
9961 1 : )]),
9962 1 : ),
9963 1 : ],
9964 1 : above_horizon: KeyLogAtLsn(vec![
9965 1 : (
9966 1 : Lsn(0x70),
9967 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
9968 1 : ),
9969 1 : (
9970 1 : Lsn(0x80),
9971 1 : Value::Image(Bytes::copy_from_slice(
9972 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
9973 1 : )),
9974 1 : ),
9975 1 : (
9976 1 : Lsn(0x90),
9977 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
9978 1 : ),
9979 1 : ]),
9980 1 : };
9981 1 : assert_eq!(res, expected_res);
9982 :
9983 : // We expect GC-compaction to run with the original GC. This would create a situation that
9984 : // the original GC algorithm removes some delta layers b/c there are full image coverage,
9985 : // therefore causing some keys to have an incomplete history below the lowest retain LSN.
9986 : // For example, we have
9987 : // ```plain
9988 : // init delta @ 0x10, image @ 0x20, delta @ 0x30 (gc_horizon), image @ 0x40.
9989 : // ```
9990 : // Now the GC horizon moves up, and we have
9991 : // ```plain
9992 : // init delta @ 0x10, image @ 0x20, delta @ 0x30, image @ 0x40 (gc_horizon)
9993 : // ```
9994 : // The original GC algorithm kicks in, and removes delta @ 0x10, image @ 0x20.
9995 : // We will end up with
9996 : // ```plain
9997 : // delta @ 0x30, image @ 0x40 (gc_horizon)
9998 : // ```
9999 : // Now we run the GC-compaction, and this key does not have a full history.
10000 : // We should be able to handle this partial history and drop everything before the
10001 : // gc_horizon image.
10002 :
10003 1 : let history = vec![
10004 1 : (
10005 1 : key,
10006 1 : Lsn(0x20),
10007 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10008 1 : ),
10009 1 : (
10010 1 : key,
10011 1 : Lsn(0x30),
10012 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
10013 1 : ),
10014 1 : (
10015 1 : key,
10016 1 : Lsn(0x40),
10017 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
10018 1 : ),
10019 1 : (
10020 1 : key,
10021 1 : Lsn(0x50),
10022 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
10023 1 : ),
10024 1 : (
10025 1 : key,
10026 1 : Lsn(0x60),
10027 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10028 1 : ),
10029 1 : (
10030 1 : key,
10031 1 : Lsn(0x70),
10032 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10033 1 : ),
10034 1 : (
10035 1 : key,
10036 1 : Lsn(0x80),
10037 1 : Value::Image(Bytes::copy_from_slice(
10038 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
10039 1 : )),
10040 1 : ),
10041 1 : (
10042 1 : key,
10043 1 : Lsn(0x90),
10044 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
10045 1 : ),
10046 : ];
10047 1 : let res = tline
10048 1 : .generate_key_retention(
10049 1 : key,
10050 1 : &history,
10051 1 : Lsn(0x60),
10052 1 : &[Lsn(0x40), Lsn(0x50)],
10053 1 : 3,
10054 1 : None,
10055 1 : true,
10056 1 : )
10057 1 : .await
10058 1 : .unwrap();
10059 1 : let expected_res = KeyHistoryRetention {
10060 1 : below_horizon: vec![
10061 1 : (
10062 1 : Lsn(0x40),
10063 1 : KeyLogAtLsn(vec![(
10064 1 : Lsn(0x40),
10065 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")),
10066 1 : )]),
10067 1 : ),
10068 1 : (
10069 1 : Lsn(0x50),
10070 1 : KeyLogAtLsn(vec![(
10071 1 : Lsn(0x50),
10072 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x50")),
10073 1 : )]),
10074 1 : ),
10075 1 : (
10076 1 : Lsn(0x60),
10077 1 : KeyLogAtLsn(vec![(
10078 1 : Lsn(0x60),
10079 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10080 1 : )]),
10081 1 : ),
10082 1 : ],
10083 1 : above_horizon: KeyLogAtLsn(vec![
10084 1 : (
10085 1 : Lsn(0x70),
10086 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10087 1 : ),
10088 1 : (
10089 1 : Lsn(0x80),
10090 1 : Value::Image(Bytes::copy_from_slice(
10091 1 : b"0x10;0x20;0x30;0x40;0x50;0x60;0x70;0x80",
10092 1 : )),
10093 1 : ),
10094 1 : (
10095 1 : Lsn(0x90),
10096 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x90")),
10097 1 : ),
10098 1 : ]),
10099 1 : };
10100 1 : assert_eq!(res, expected_res);
10101 :
10102 : // In case of branch compaction, the branch itself does not have the full history, and we need to provide
10103 : // the ancestor image in the test case.
10104 :
10105 1 : let history = vec![
10106 1 : (
10107 1 : key,
10108 1 : Lsn(0x20),
10109 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10110 1 : ),
10111 1 : (
10112 1 : key,
10113 1 : Lsn(0x30),
10114 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x30")),
10115 1 : ),
10116 1 : (
10117 1 : key,
10118 1 : Lsn(0x40),
10119 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
10120 1 : ),
10121 1 : (
10122 1 : key,
10123 1 : Lsn(0x70),
10124 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10125 1 : ),
10126 : ];
10127 1 : let res = tline
10128 1 : .generate_key_retention(
10129 1 : key,
10130 1 : &history,
10131 1 : Lsn(0x60),
10132 1 : &[],
10133 1 : 3,
10134 1 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
10135 1 : true,
10136 1 : )
10137 1 : .await
10138 1 : .unwrap();
10139 1 : let expected_res = KeyHistoryRetention {
10140 1 : below_horizon: vec![(
10141 1 : Lsn(0x60),
10142 1 : KeyLogAtLsn(vec![(
10143 1 : Lsn(0x60),
10144 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x30;0x40")), // use the ancestor image to reconstruct the page
10145 1 : )]),
10146 1 : )],
10147 1 : above_horizon: KeyLogAtLsn(vec![(
10148 1 : Lsn(0x70),
10149 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10150 1 : )]),
10151 1 : };
10152 1 : assert_eq!(res, expected_res);
10153 :
10154 1 : let history = vec![
10155 1 : (
10156 1 : key,
10157 1 : Lsn(0x20),
10158 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10159 1 : ),
10160 1 : (
10161 1 : key,
10162 1 : Lsn(0x40),
10163 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x40")),
10164 1 : ),
10165 1 : (
10166 1 : key,
10167 1 : Lsn(0x60),
10168 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x60")),
10169 1 : ),
10170 1 : (
10171 1 : key,
10172 1 : Lsn(0x70),
10173 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10174 1 : ),
10175 : ];
10176 1 : let res = tline
10177 1 : .generate_key_retention(
10178 1 : key,
10179 1 : &history,
10180 1 : Lsn(0x60),
10181 1 : &[Lsn(0x30)],
10182 1 : 3,
10183 1 : Some((key, Lsn(0x10), Bytes::copy_from_slice(b"0x10"))),
10184 1 : true,
10185 1 : )
10186 1 : .await
10187 1 : .unwrap();
10188 1 : let expected_res = KeyHistoryRetention {
10189 1 : below_horizon: vec![
10190 1 : (
10191 1 : Lsn(0x30),
10192 1 : KeyLogAtLsn(vec![(
10193 1 : Lsn(0x20),
10194 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x20")),
10195 1 : )]),
10196 1 : ),
10197 1 : (
10198 1 : Lsn(0x60),
10199 1 : KeyLogAtLsn(vec![(
10200 1 : Lsn(0x60),
10201 1 : Value::Image(Bytes::copy_from_slice(b"0x10;0x20;0x40;0x60")),
10202 1 : )]),
10203 1 : ),
10204 1 : ],
10205 1 : above_horizon: KeyLogAtLsn(vec![(
10206 1 : Lsn(0x70),
10207 1 : Value::WalRecord(NeonWalRecord::wal_append(";0x70")),
10208 1 : )]),
10209 1 : };
10210 1 : assert_eq!(res, expected_res);
10211 :
10212 2 : Ok(())
10213 1 : }
10214 :
10215 : #[cfg(feature = "testing")]
10216 : #[tokio::test]
10217 1 : async fn test_simple_bottom_most_compaction_with_retain_lsns() -> anyhow::Result<()> {
10218 1 : let harness =
10219 1 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns").await?;
10220 1 : let (tenant, ctx) = harness.load().await;
10221 :
10222 259 : fn get_key(id: u32) -> Key {
10223 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10224 259 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10225 259 : key.field6 = id;
10226 259 : key
10227 259 : }
10228 :
10229 1 : let img_layer = (0..10)
10230 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10231 1 : .collect_vec();
10232 :
10233 1 : let delta1 = vec![
10234 1 : (
10235 1 : get_key(1),
10236 1 : Lsn(0x20),
10237 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10238 1 : ),
10239 1 : (
10240 1 : get_key(2),
10241 1 : Lsn(0x30),
10242 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10243 1 : ),
10244 1 : (
10245 1 : get_key(3),
10246 1 : Lsn(0x28),
10247 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10248 1 : ),
10249 1 : (
10250 1 : get_key(3),
10251 1 : Lsn(0x30),
10252 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10253 1 : ),
10254 1 : (
10255 1 : get_key(3),
10256 1 : Lsn(0x40),
10257 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
10258 1 : ),
10259 : ];
10260 1 : let delta2 = vec![
10261 1 : (
10262 1 : get_key(5),
10263 1 : Lsn(0x20),
10264 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10265 1 : ),
10266 1 : (
10267 1 : get_key(6),
10268 1 : Lsn(0x20),
10269 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10270 1 : ),
10271 : ];
10272 1 : let delta3 = vec![
10273 1 : (
10274 1 : get_key(8),
10275 1 : Lsn(0x48),
10276 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10277 1 : ),
10278 1 : (
10279 1 : get_key(9),
10280 1 : Lsn(0x48),
10281 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10282 1 : ),
10283 : ];
10284 :
10285 1 : let tline = tenant
10286 1 : .create_test_timeline_with_layers(
10287 1 : TIMELINE_ID,
10288 1 : Lsn(0x10),
10289 1 : DEFAULT_PG_VERSION,
10290 1 : &ctx,
10291 1 : Vec::new(), // in-memory layers
10292 1 : vec![
10293 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
10294 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
10295 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10296 1 : ], // delta layers
10297 1 : vec![(Lsn(0x10), img_layer)], // image layers
10298 1 : Lsn(0x50),
10299 1 : )
10300 1 : .await?;
10301 : {
10302 1 : tline
10303 1 : .applied_gc_cutoff_lsn
10304 1 : .lock_for_write()
10305 1 : .store_and_unlock(Lsn(0x30))
10306 1 : .wait()
10307 1 : .await;
10308 : // Update GC info
10309 1 : let mut guard = tline.gc_info.write().unwrap();
10310 1 : *guard = GcInfo {
10311 1 : retain_lsns: vec![
10312 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10313 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10314 1 : ],
10315 1 : cutoffs: GcCutoffs {
10316 1 : time: Some(Lsn(0x30)),
10317 1 : space: Lsn(0x30),
10318 1 : },
10319 1 : leases: Default::default(),
10320 1 : within_ancestor_pitr: false,
10321 1 : };
10322 : }
10323 :
10324 1 : let expected_result = [
10325 1 : Bytes::from_static(b"value 0@0x10"),
10326 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10327 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10328 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10329 1 : Bytes::from_static(b"value 4@0x10"),
10330 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10331 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10332 1 : Bytes::from_static(b"value 7@0x10"),
10333 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10334 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10335 1 : ];
10336 :
10337 1 : let expected_result_at_gc_horizon = [
10338 1 : Bytes::from_static(b"value 0@0x10"),
10339 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10340 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10341 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
10342 1 : Bytes::from_static(b"value 4@0x10"),
10343 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10344 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10345 1 : Bytes::from_static(b"value 7@0x10"),
10346 1 : Bytes::from_static(b"value 8@0x10"),
10347 1 : Bytes::from_static(b"value 9@0x10"),
10348 1 : ];
10349 :
10350 1 : let expected_result_at_lsn_20 = [
10351 1 : Bytes::from_static(b"value 0@0x10"),
10352 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10353 1 : Bytes::from_static(b"value 2@0x10"),
10354 1 : Bytes::from_static(b"value 3@0x10"),
10355 1 : Bytes::from_static(b"value 4@0x10"),
10356 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10357 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10358 1 : Bytes::from_static(b"value 7@0x10"),
10359 1 : Bytes::from_static(b"value 8@0x10"),
10360 1 : Bytes::from_static(b"value 9@0x10"),
10361 1 : ];
10362 :
10363 1 : let expected_result_at_lsn_10 = [
10364 1 : Bytes::from_static(b"value 0@0x10"),
10365 1 : Bytes::from_static(b"value 1@0x10"),
10366 1 : Bytes::from_static(b"value 2@0x10"),
10367 1 : Bytes::from_static(b"value 3@0x10"),
10368 1 : Bytes::from_static(b"value 4@0x10"),
10369 1 : Bytes::from_static(b"value 5@0x10"),
10370 1 : Bytes::from_static(b"value 6@0x10"),
10371 1 : Bytes::from_static(b"value 7@0x10"),
10372 1 : Bytes::from_static(b"value 8@0x10"),
10373 1 : Bytes::from_static(b"value 9@0x10"),
10374 1 : ];
10375 :
10376 6 : let verify_result = || async {
10377 6 : let gc_horizon = {
10378 6 : let gc_info = tline.gc_info.read().unwrap();
10379 6 : gc_info.cutoffs.time.unwrap_or_default()
10380 : };
10381 66 : for idx in 0..10 {
10382 60 : assert_eq!(
10383 60 : tline
10384 60 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10385 60 : .await
10386 60 : .unwrap(),
10387 60 : &expected_result[idx]
10388 : );
10389 60 : assert_eq!(
10390 60 : tline
10391 60 : .get(get_key(idx as u32), gc_horizon, &ctx)
10392 60 : .await
10393 60 : .unwrap(),
10394 60 : &expected_result_at_gc_horizon[idx]
10395 : );
10396 60 : assert_eq!(
10397 60 : tline
10398 60 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10399 60 : .await
10400 60 : .unwrap(),
10401 60 : &expected_result_at_lsn_20[idx]
10402 : );
10403 60 : assert_eq!(
10404 60 : tline
10405 60 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10406 60 : .await
10407 60 : .unwrap(),
10408 60 : &expected_result_at_lsn_10[idx]
10409 : );
10410 : }
10411 12 : };
10412 :
10413 1 : verify_result().await;
10414 :
10415 1 : let cancel = CancellationToken::new();
10416 1 : let mut dryrun_flags = EnumSet::new();
10417 1 : dryrun_flags.insert(CompactFlags::DryRun);
10418 :
10419 1 : tline
10420 1 : .compact_with_gc(
10421 1 : &cancel,
10422 1 : CompactOptions {
10423 1 : flags: dryrun_flags,
10424 1 : ..Default::default()
10425 1 : },
10426 1 : &ctx,
10427 1 : )
10428 1 : .await
10429 1 : .unwrap();
10430 : // We expect layer map to be the same b/c the dry run flag, but we don't know whether there will be other background jobs
10431 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10432 1 : verify_result().await;
10433 :
10434 1 : tline
10435 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10436 1 : .await
10437 1 : .unwrap();
10438 1 : verify_result().await;
10439 :
10440 : // compact again
10441 1 : tline
10442 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10443 1 : .await
10444 1 : .unwrap();
10445 1 : verify_result().await;
10446 :
10447 : // increase GC horizon and compact again
10448 : {
10449 1 : tline
10450 1 : .applied_gc_cutoff_lsn
10451 1 : .lock_for_write()
10452 1 : .store_and_unlock(Lsn(0x38))
10453 1 : .wait()
10454 1 : .await;
10455 : // Update GC info
10456 1 : let mut guard = tline.gc_info.write().unwrap();
10457 1 : guard.cutoffs.time = Some(Lsn(0x38));
10458 1 : guard.cutoffs.space = Lsn(0x38);
10459 : }
10460 1 : tline
10461 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10462 1 : .await
10463 1 : .unwrap();
10464 1 : verify_result().await; // no wals between 0x30 and 0x38, so we should obtain the same result
10465 :
10466 : // not increasing the GC horizon and compact again
10467 1 : tline
10468 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10469 1 : .await
10470 1 : .unwrap();
10471 1 : verify_result().await;
10472 :
10473 2 : Ok(())
10474 1 : }
10475 :
10476 : #[cfg(feature = "testing")]
10477 : #[tokio::test]
10478 1 : async fn test_simple_bottom_most_compaction_with_retain_lsns_single_key() -> anyhow::Result<()>
10479 1 : {
10480 1 : let harness =
10481 1 : TenantHarness::create("test_simple_bottom_most_compaction_with_retain_lsns_single_key")
10482 1 : .await?;
10483 1 : let (tenant, ctx) = harness.load().await;
10484 :
10485 176 : fn get_key(id: u32) -> Key {
10486 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
10487 176 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
10488 176 : key.field6 = id;
10489 176 : key
10490 176 : }
10491 :
10492 1 : let img_layer = (0..10)
10493 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10494 1 : .collect_vec();
10495 :
10496 1 : let delta1 = vec![
10497 1 : (
10498 1 : get_key(1),
10499 1 : Lsn(0x20),
10500 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10501 1 : ),
10502 1 : (
10503 1 : get_key(1),
10504 1 : Lsn(0x28),
10505 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10506 1 : ),
10507 : ];
10508 1 : let delta2 = vec![
10509 1 : (
10510 1 : get_key(1),
10511 1 : Lsn(0x30),
10512 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10513 1 : ),
10514 1 : (
10515 1 : get_key(1),
10516 1 : Lsn(0x38),
10517 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
10518 1 : ),
10519 : ];
10520 1 : let delta3 = vec![
10521 1 : (
10522 1 : get_key(8),
10523 1 : Lsn(0x48),
10524 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10525 1 : ),
10526 1 : (
10527 1 : get_key(9),
10528 1 : Lsn(0x48),
10529 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10530 1 : ),
10531 : ];
10532 :
10533 1 : let tline = tenant
10534 1 : .create_test_timeline_with_layers(
10535 1 : TIMELINE_ID,
10536 1 : Lsn(0x10),
10537 1 : DEFAULT_PG_VERSION,
10538 1 : &ctx,
10539 1 : Vec::new(), // in-memory layers
10540 1 : vec![
10541 1 : // delta1 and delta 2 only contain a single key but multiple updates
10542 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x30), delta1),
10543 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
10544 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x50), delta3),
10545 1 : ], // delta layers
10546 1 : vec![(Lsn(0x10), img_layer)], // image layers
10547 1 : Lsn(0x50),
10548 1 : )
10549 1 : .await?;
10550 : {
10551 1 : tline
10552 1 : .applied_gc_cutoff_lsn
10553 1 : .lock_for_write()
10554 1 : .store_and_unlock(Lsn(0x30))
10555 1 : .wait()
10556 1 : .await;
10557 : // Update GC info
10558 1 : let mut guard = tline.gc_info.write().unwrap();
10559 1 : *guard = GcInfo {
10560 1 : retain_lsns: vec![
10561 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
10562 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
10563 1 : ],
10564 1 : cutoffs: GcCutoffs {
10565 1 : time: Some(Lsn(0x30)),
10566 1 : space: Lsn(0x30),
10567 1 : },
10568 1 : leases: Default::default(),
10569 1 : within_ancestor_pitr: false,
10570 1 : };
10571 : }
10572 :
10573 1 : let expected_result = [
10574 1 : Bytes::from_static(b"value 0@0x10"),
10575 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
10576 1 : Bytes::from_static(b"value 2@0x10"),
10577 1 : Bytes::from_static(b"value 3@0x10"),
10578 1 : Bytes::from_static(b"value 4@0x10"),
10579 1 : Bytes::from_static(b"value 5@0x10"),
10580 1 : Bytes::from_static(b"value 6@0x10"),
10581 1 : Bytes::from_static(b"value 7@0x10"),
10582 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10583 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10584 1 : ];
10585 :
10586 1 : let expected_result_at_gc_horizon = [
10587 1 : Bytes::from_static(b"value 0@0x10"),
10588 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
10589 1 : Bytes::from_static(b"value 2@0x10"),
10590 1 : Bytes::from_static(b"value 3@0x10"),
10591 1 : Bytes::from_static(b"value 4@0x10"),
10592 1 : Bytes::from_static(b"value 5@0x10"),
10593 1 : Bytes::from_static(b"value 6@0x10"),
10594 1 : Bytes::from_static(b"value 7@0x10"),
10595 1 : Bytes::from_static(b"value 8@0x10"),
10596 1 : Bytes::from_static(b"value 9@0x10"),
10597 1 : ];
10598 :
10599 1 : let expected_result_at_lsn_20 = [
10600 1 : Bytes::from_static(b"value 0@0x10"),
10601 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10602 1 : Bytes::from_static(b"value 2@0x10"),
10603 1 : Bytes::from_static(b"value 3@0x10"),
10604 1 : Bytes::from_static(b"value 4@0x10"),
10605 1 : Bytes::from_static(b"value 5@0x10"),
10606 1 : Bytes::from_static(b"value 6@0x10"),
10607 1 : Bytes::from_static(b"value 7@0x10"),
10608 1 : Bytes::from_static(b"value 8@0x10"),
10609 1 : Bytes::from_static(b"value 9@0x10"),
10610 1 : ];
10611 :
10612 1 : let expected_result_at_lsn_10 = [
10613 1 : Bytes::from_static(b"value 0@0x10"),
10614 1 : Bytes::from_static(b"value 1@0x10"),
10615 1 : Bytes::from_static(b"value 2@0x10"),
10616 1 : Bytes::from_static(b"value 3@0x10"),
10617 1 : Bytes::from_static(b"value 4@0x10"),
10618 1 : Bytes::from_static(b"value 5@0x10"),
10619 1 : Bytes::from_static(b"value 6@0x10"),
10620 1 : Bytes::from_static(b"value 7@0x10"),
10621 1 : Bytes::from_static(b"value 8@0x10"),
10622 1 : Bytes::from_static(b"value 9@0x10"),
10623 1 : ];
10624 :
10625 4 : let verify_result = || async {
10626 4 : let gc_horizon = {
10627 4 : let gc_info = tline.gc_info.read().unwrap();
10628 4 : gc_info.cutoffs.time.unwrap_or_default()
10629 : };
10630 44 : for idx in 0..10 {
10631 40 : assert_eq!(
10632 40 : tline
10633 40 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10634 40 : .await
10635 40 : .unwrap(),
10636 40 : &expected_result[idx]
10637 : );
10638 40 : assert_eq!(
10639 40 : tline
10640 40 : .get(get_key(idx as u32), gc_horizon, &ctx)
10641 40 : .await
10642 40 : .unwrap(),
10643 40 : &expected_result_at_gc_horizon[idx]
10644 : );
10645 40 : assert_eq!(
10646 40 : tline
10647 40 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
10648 40 : .await
10649 40 : .unwrap(),
10650 40 : &expected_result_at_lsn_20[idx]
10651 : );
10652 40 : assert_eq!(
10653 40 : tline
10654 40 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
10655 40 : .await
10656 40 : .unwrap(),
10657 40 : &expected_result_at_lsn_10[idx]
10658 : );
10659 : }
10660 8 : };
10661 :
10662 1 : verify_result().await;
10663 :
10664 1 : let cancel = CancellationToken::new();
10665 1 : let mut dryrun_flags = EnumSet::new();
10666 1 : dryrun_flags.insert(CompactFlags::DryRun);
10667 :
10668 1 : tline
10669 1 : .compact_with_gc(
10670 1 : &cancel,
10671 1 : CompactOptions {
10672 1 : flags: dryrun_flags,
10673 1 : ..Default::default()
10674 1 : },
10675 1 : &ctx,
10676 1 : )
10677 1 : .await
10678 1 : .unwrap();
10679 : // We expect layer map to be the same b/c the dry run flag, but we don't know whether there will be other background jobs
10680 : // cleaning things up, and therefore, we don't do sanity checks on the layer map during unit tests.
10681 1 : verify_result().await;
10682 :
10683 1 : tline
10684 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10685 1 : .await
10686 1 : .unwrap();
10687 1 : verify_result().await;
10688 :
10689 : // compact again
10690 1 : tline
10691 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10692 1 : .await
10693 1 : .unwrap();
10694 1 : verify_result().await;
10695 :
10696 2 : Ok(())
10697 1 : }
10698 :
10699 : #[cfg(feature = "testing")]
10700 : #[tokio::test]
10701 1 : async fn test_simple_bottom_most_compaction_on_branch() -> anyhow::Result<()> {
10702 : use models::CompactLsnRange;
10703 :
10704 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_on_branch").await?;
10705 1 : let (tenant, ctx) = harness.load().await;
10706 :
10707 83 : fn get_key(id: u32) -> Key {
10708 83 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
10709 83 : key.field6 = id;
10710 83 : key
10711 83 : }
10712 :
10713 1 : let img_layer = (0..10)
10714 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
10715 1 : .collect_vec();
10716 :
10717 1 : let delta1 = vec![
10718 1 : (
10719 1 : get_key(1),
10720 1 : Lsn(0x20),
10721 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10722 1 : ),
10723 1 : (
10724 1 : get_key(2),
10725 1 : Lsn(0x30),
10726 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10727 1 : ),
10728 1 : (
10729 1 : get_key(3),
10730 1 : Lsn(0x28),
10731 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
10732 1 : ),
10733 1 : (
10734 1 : get_key(3),
10735 1 : Lsn(0x30),
10736 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
10737 1 : ),
10738 1 : (
10739 1 : get_key(3),
10740 1 : Lsn(0x40),
10741 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
10742 1 : ),
10743 : ];
10744 1 : let delta2 = vec![
10745 1 : (
10746 1 : get_key(5),
10747 1 : Lsn(0x20),
10748 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10749 1 : ),
10750 1 : (
10751 1 : get_key(6),
10752 1 : Lsn(0x20),
10753 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
10754 1 : ),
10755 : ];
10756 1 : let delta3 = vec![
10757 1 : (
10758 1 : get_key(8),
10759 1 : Lsn(0x48),
10760 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10761 1 : ),
10762 1 : (
10763 1 : get_key(9),
10764 1 : Lsn(0x48),
10765 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
10766 1 : ),
10767 : ];
10768 :
10769 1 : let parent_tline = tenant
10770 1 : .create_test_timeline_with_layers(
10771 1 : TIMELINE_ID,
10772 1 : Lsn(0x10),
10773 1 : DEFAULT_PG_VERSION,
10774 1 : &ctx,
10775 1 : vec![], // in-memory layers
10776 1 : vec![], // delta layers
10777 1 : vec![(Lsn(0x18), img_layer)], // image layers
10778 1 : Lsn(0x18),
10779 1 : )
10780 1 : .await?;
10781 :
10782 1 : parent_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10783 :
10784 1 : let branch_tline = tenant
10785 1 : .branch_timeline_test_with_layers(
10786 1 : &parent_tline,
10787 1 : NEW_TIMELINE_ID,
10788 1 : Some(Lsn(0x18)),
10789 1 : &ctx,
10790 1 : vec![
10791 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
10792 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
10793 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
10794 1 : ], // delta layers
10795 1 : vec![], // image layers
10796 1 : Lsn(0x50),
10797 1 : )
10798 1 : .await?;
10799 :
10800 1 : branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
10801 :
10802 : {
10803 1 : parent_tline
10804 1 : .applied_gc_cutoff_lsn
10805 1 : .lock_for_write()
10806 1 : .store_and_unlock(Lsn(0x10))
10807 1 : .wait()
10808 1 : .await;
10809 : // Update GC info
10810 1 : let mut guard = parent_tline.gc_info.write().unwrap();
10811 1 : *guard = GcInfo {
10812 1 : retain_lsns: vec![(Lsn(0x18), branch_tline.timeline_id, MaybeOffloaded::No)],
10813 1 : cutoffs: GcCutoffs {
10814 1 : time: Some(Lsn(0x10)),
10815 1 : space: Lsn(0x10),
10816 1 : },
10817 1 : leases: Default::default(),
10818 1 : within_ancestor_pitr: false,
10819 1 : };
10820 : }
10821 :
10822 : {
10823 1 : branch_tline
10824 1 : .applied_gc_cutoff_lsn
10825 1 : .lock_for_write()
10826 1 : .store_and_unlock(Lsn(0x50))
10827 1 : .wait()
10828 1 : .await;
10829 : // Update GC info
10830 1 : let mut guard = branch_tline.gc_info.write().unwrap();
10831 1 : *guard = GcInfo {
10832 1 : retain_lsns: vec![(Lsn(0x40), branch_tline.timeline_id, MaybeOffloaded::No)],
10833 1 : cutoffs: GcCutoffs {
10834 1 : time: Some(Lsn(0x50)),
10835 1 : space: Lsn(0x50),
10836 1 : },
10837 1 : leases: Default::default(),
10838 1 : within_ancestor_pitr: false,
10839 1 : };
10840 : }
10841 :
10842 1 : let expected_result_at_gc_horizon = [
10843 1 : Bytes::from_static(b"value 0@0x10"),
10844 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10845 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10846 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10847 1 : Bytes::from_static(b"value 4@0x10"),
10848 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10849 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10850 1 : Bytes::from_static(b"value 7@0x10"),
10851 1 : Bytes::from_static(b"value 8@0x10@0x48"),
10852 1 : Bytes::from_static(b"value 9@0x10@0x48"),
10853 1 : ];
10854 :
10855 1 : let expected_result_at_lsn_40 = [
10856 1 : Bytes::from_static(b"value 0@0x10"),
10857 1 : Bytes::from_static(b"value 1@0x10@0x20"),
10858 1 : Bytes::from_static(b"value 2@0x10@0x30"),
10859 1 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
10860 1 : Bytes::from_static(b"value 4@0x10"),
10861 1 : Bytes::from_static(b"value 5@0x10@0x20"),
10862 1 : Bytes::from_static(b"value 6@0x10@0x20"),
10863 1 : Bytes::from_static(b"value 7@0x10"),
10864 1 : Bytes::from_static(b"value 8@0x10"),
10865 1 : Bytes::from_static(b"value 9@0x10"),
10866 1 : ];
10867 :
10868 3 : let verify_result = || async {
10869 33 : for idx in 0..10 {
10870 30 : assert_eq!(
10871 30 : branch_tline
10872 30 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
10873 30 : .await
10874 30 : .unwrap(),
10875 30 : &expected_result_at_gc_horizon[idx]
10876 : );
10877 30 : assert_eq!(
10878 30 : branch_tline
10879 30 : .get(get_key(idx as u32), Lsn(0x40), &ctx)
10880 30 : .await
10881 30 : .unwrap(),
10882 30 : &expected_result_at_lsn_40[idx]
10883 : );
10884 : }
10885 6 : };
10886 :
10887 1 : verify_result().await;
10888 :
10889 1 : let cancel = CancellationToken::new();
10890 1 : branch_tline
10891 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
10892 1 : .await
10893 1 : .unwrap();
10894 :
10895 1 : verify_result().await;
10896 :
10897 : // Piggyback a compaction with above_lsn. Ensure it works correctly when the specified LSN intersects with the layer files.
10898 : // Now we already have a single large delta layer, so the compaction min_layer_lsn should be the same as ancestor LSN (0x18).
10899 1 : branch_tline
10900 1 : .compact_with_gc(
10901 1 : &cancel,
10902 1 : CompactOptions {
10903 1 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x40))),
10904 1 : ..Default::default()
10905 1 : },
10906 1 : &ctx,
10907 1 : )
10908 1 : .await
10909 1 : .unwrap();
10910 :
10911 1 : verify_result().await;
10912 :
10913 2 : Ok(())
10914 1 : }
10915 :
10916 : // Regression test for https://github.com/neondatabase/neon/issues/9012
10917 : // Create an image arrangement where we have to read at different LSN ranges
10918 : // from a delta layer. This is achieved by overlapping an image layer on top of
10919 : // a delta layer. Like so:
10920 : //
10921 : // A B
10922 : // +----------------+ -> delta_layer
10923 : // | | ^ lsn
10924 : // | =========|-> nested_image_layer |
10925 : // | C | |
10926 : // +----------------+ |
10927 : // ======== -> baseline_image_layer +-------> key
10928 : //
10929 : //
10930 : // When querying the key range [A, B) we need to read at different LSN ranges
10931 : // for [A, C) and [C, B). This test checks that the described edge case is handled correctly.
10932 : #[cfg(feature = "testing")]
10933 : #[tokio::test]
10934 1 : async fn test_vectored_read_with_nested_image_layer() -> anyhow::Result<()> {
10935 1 : let harness = TenantHarness::create("test_vectored_read_with_nested_image_layer").await?;
10936 1 : let (tenant, ctx) = harness.load().await;
10937 :
10938 1 : let will_init_keys = [2, 6];
10939 22 : fn get_key(id: u32) -> Key {
10940 22 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
10941 22 : key.field6 = id;
10942 22 : key
10943 22 : }
10944 :
10945 1 : let mut expected_key_values = HashMap::new();
10946 :
10947 1 : let baseline_image_layer_lsn = Lsn(0x10);
10948 1 : let mut baseline_img_layer = Vec::new();
10949 6 : for i in 0..5 {
10950 5 : let key = get_key(i);
10951 5 : let value = format!("value {i}@{baseline_image_layer_lsn}");
10952 :
10953 5 : let removed = expected_key_values.insert(key, value.clone());
10954 5 : assert!(removed.is_none());
10955 :
10956 5 : baseline_img_layer.push((key, Bytes::from(value)));
10957 : }
10958 :
10959 1 : let nested_image_layer_lsn = Lsn(0x50);
10960 1 : let mut nested_img_layer = Vec::new();
10961 6 : for i in 5..10 {
10962 5 : let key = get_key(i);
10963 5 : let value = format!("value {i}@{nested_image_layer_lsn}");
10964 :
10965 5 : let removed = expected_key_values.insert(key, value.clone());
10966 5 : assert!(removed.is_none());
10967 :
10968 5 : nested_img_layer.push((key, Bytes::from(value)));
10969 : }
10970 :
10971 1 : let mut delta_layer_spec = Vec::default();
10972 1 : let delta_layer_start_lsn = Lsn(0x20);
10973 1 : let mut delta_layer_end_lsn = delta_layer_start_lsn;
10974 :
10975 11 : for i in 0..10 {
10976 10 : let key = get_key(i);
10977 10 : let key_in_nested = nested_img_layer
10978 10 : .iter()
10979 40 : .any(|(key_with_img, _)| *key_with_img == key);
10980 10 : let lsn = {
10981 10 : if key_in_nested {
10982 5 : Lsn(nested_image_layer_lsn.0 + 0x10)
10983 : } else {
10984 5 : delta_layer_start_lsn
10985 : }
10986 : };
10987 :
10988 10 : let will_init = will_init_keys.contains(&i);
10989 10 : if will_init {
10990 2 : delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
10991 2 :
10992 2 : expected_key_values.insert(key, "".to_string());
10993 8 : } else {
10994 8 : let delta = format!("@{lsn}");
10995 8 : delta_layer_spec.push((
10996 8 : key,
10997 8 : lsn,
10998 8 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
10999 8 : ));
11000 8 :
11001 8 : expected_key_values
11002 8 : .get_mut(&key)
11003 8 : .expect("An image exists for each key")
11004 8 : .push_str(delta.as_str());
11005 8 : }
11006 10 : delta_layer_end_lsn = std::cmp::max(delta_layer_start_lsn, lsn);
11007 : }
11008 :
11009 1 : delta_layer_end_lsn = Lsn(delta_layer_end_lsn.0 + 1);
11010 :
11011 1 : assert!(
11012 1 : nested_image_layer_lsn > delta_layer_start_lsn
11013 1 : && nested_image_layer_lsn < delta_layer_end_lsn
11014 : );
11015 :
11016 1 : let tline = tenant
11017 1 : .create_test_timeline_with_layers(
11018 1 : TIMELINE_ID,
11019 1 : baseline_image_layer_lsn,
11020 1 : DEFAULT_PG_VERSION,
11021 1 : &ctx,
11022 1 : vec![], // in-memory layers
11023 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
11024 1 : delta_layer_start_lsn..delta_layer_end_lsn,
11025 1 : delta_layer_spec,
11026 1 : )], // delta layers
11027 1 : vec![
11028 1 : (baseline_image_layer_lsn, baseline_img_layer),
11029 1 : (nested_image_layer_lsn, nested_img_layer),
11030 1 : ], // image layers
11031 1 : delta_layer_end_lsn,
11032 1 : )
11033 1 : .await?;
11034 :
11035 1 : let query = VersionedKeySpaceQuery::uniform(
11036 1 : KeySpace::single(get_key(0)..get_key(10)),
11037 1 : delta_layer_end_lsn,
11038 : );
11039 :
11040 1 : let results = tline
11041 1 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
11042 1 : .await
11043 1 : .expect("No vectored errors");
11044 11 : for (key, res) in results {
11045 10 : let value = res.expect("No key errors");
11046 10 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
11047 10 : assert_eq!(value, Bytes::from(expected_value));
11048 1 : }
11049 1 :
11050 1 : Ok(())
11051 1 : }
11052 :
11053 : #[cfg(feature = "testing")]
11054 : #[tokio::test]
11055 1 : async fn test_vectored_read_with_image_layer_inside_inmem() -> anyhow::Result<()> {
11056 1 : let harness =
11057 1 : TenantHarness::create("test_vectored_read_with_image_layer_inside_inmem").await?;
11058 1 : let (tenant, ctx) = harness.load().await;
11059 :
11060 1 : let will_init_keys = [2, 6];
11061 32 : fn get_key(id: u32) -> Key {
11062 32 : let mut key = Key::from_hex("110000000033333333444444445500000000").unwrap();
11063 32 : key.field6 = id;
11064 32 : key
11065 32 : }
11066 :
11067 1 : let mut expected_key_values = HashMap::new();
11068 :
11069 1 : let baseline_image_layer_lsn = Lsn(0x10);
11070 1 : let mut baseline_img_layer = Vec::new();
11071 6 : for i in 0..5 {
11072 5 : let key = get_key(i);
11073 5 : let value = format!("value {i}@{baseline_image_layer_lsn}");
11074 :
11075 5 : let removed = expected_key_values.insert(key, value.clone());
11076 5 : assert!(removed.is_none());
11077 :
11078 5 : baseline_img_layer.push((key, Bytes::from(value)));
11079 : }
11080 :
11081 1 : let nested_image_layer_lsn = Lsn(0x50);
11082 1 : let mut nested_img_layer = Vec::new();
11083 6 : for i in 5..10 {
11084 5 : let key = get_key(i);
11085 5 : let value = format!("value {i}@{nested_image_layer_lsn}");
11086 :
11087 5 : let removed = expected_key_values.insert(key, value.clone());
11088 5 : assert!(removed.is_none());
11089 :
11090 5 : nested_img_layer.push((key, Bytes::from(value)));
11091 : }
11092 :
11093 1 : let frozen_layer = {
11094 1 : let lsn_range = Lsn(0x40)..Lsn(0x60);
11095 1 : let mut data = Vec::new();
11096 11 : for i in 0..10 {
11097 10 : let key = get_key(i);
11098 10 : let key_in_nested = nested_img_layer
11099 10 : .iter()
11100 40 : .any(|(key_with_img, _)| *key_with_img == key);
11101 10 : let lsn = {
11102 10 : if key_in_nested {
11103 5 : Lsn(nested_image_layer_lsn.0 + 5)
11104 : } else {
11105 5 : lsn_range.start
11106 : }
11107 : };
11108 :
11109 10 : let will_init = will_init_keys.contains(&i);
11110 10 : if will_init {
11111 2 : data.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
11112 2 :
11113 2 : expected_key_values.insert(key, "".to_string());
11114 8 : } else {
11115 8 : let delta = format!("@{lsn}");
11116 8 : data.push((
11117 8 : key,
11118 8 : lsn,
11119 8 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
11120 8 : ));
11121 8 :
11122 8 : expected_key_values
11123 8 : .get_mut(&key)
11124 8 : .expect("An image exists for each key")
11125 8 : .push_str(delta.as_str());
11126 8 : }
11127 : }
11128 :
11129 1 : InMemoryLayerTestDesc {
11130 1 : lsn_range,
11131 1 : is_open: false,
11132 1 : data,
11133 1 : }
11134 : };
11135 :
11136 1 : let (open_layer, last_record_lsn) = {
11137 1 : let start_lsn = Lsn(0x70);
11138 1 : let mut data = Vec::new();
11139 1 : let mut end_lsn = Lsn(0);
11140 11 : for i in 0..10 {
11141 10 : let key = get_key(i);
11142 10 : let lsn = Lsn(start_lsn.0 + i as u64);
11143 10 : let delta = format!("@{lsn}");
11144 10 : data.push((
11145 10 : key,
11146 10 : lsn,
11147 10 : Value::WalRecord(NeonWalRecord::wal_append(&delta)),
11148 10 : ));
11149 10 :
11150 10 : expected_key_values
11151 10 : .get_mut(&key)
11152 10 : .expect("An image exists for each key")
11153 10 : .push_str(delta.as_str());
11154 10 :
11155 10 : end_lsn = std::cmp::max(end_lsn, lsn);
11156 10 : }
11157 :
11158 1 : (
11159 1 : InMemoryLayerTestDesc {
11160 1 : lsn_range: start_lsn..Lsn::MAX,
11161 1 : is_open: true,
11162 1 : data,
11163 1 : },
11164 1 : end_lsn,
11165 1 : )
11166 : };
11167 :
11168 1 : assert!(
11169 1 : nested_image_layer_lsn > frozen_layer.lsn_range.start
11170 1 : && nested_image_layer_lsn < frozen_layer.lsn_range.end
11171 : );
11172 :
11173 1 : let tline = tenant
11174 1 : .create_test_timeline_with_layers(
11175 1 : TIMELINE_ID,
11176 1 : baseline_image_layer_lsn,
11177 1 : DEFAULT_PG_VERSION,
11178 1 : &ctx,
11179 1 : vec![open_layer, frozen_layer], // in-memory layers
11180 1 : Vec::new(), // delta layers
11181 1 : vec![
11182 1 : (baseline_image_layer_lsn, baseline_img_layer),
11183 1 : (nested_image_layer_lsn, nested_img_layer),
11184 1 : ], // image layers
11185 1 : last_record_lsn,
11186 1 : )
11187 1 : .await?;
11188 :
11189 1 : let query = VersionedKeySpaceQuery::uniform(
11190 1 : KeySpace::single(get_key(0)..get_key(10)),
11191 1 : last_record_lsn,
11192 : );
11193 :
11194 1 : let results = tline
11195 1 : .get_vectored(query, IoConcurrency::sequential(), &ctx)
11196 1 : .await
11197 1 : .expect("No vectored errors");
11198 11 : for (key, res) in results {
11199 10 : let value = res.expect("No key errors");
11200 10 : let expected_value = expected_key_values.remove(&key).expect("No unknown keys");
11201 10 : assert_eq!(value, Bytes::from(expected_value.clone()));
11202 1 :
11203 10 : tracing::info!("key={key} value={expected_value}");
11204 1 : }
11205 1 :
11206 1 : Ok(())
11207 1 : }
11208 :
11209 : // A randomized read path test. Generates a layer map according to a deterministic
11210 : // specification. Fills the (key, LSN) space in random manner and then performs
11211 : // random scattered queries validating the results against in-memory storage.
11212 : //
11213 : // See this internal Notion page for a diagram of the layer map:
11214 : // https://www.notion.so/neondatabase/Read-Path-Unit-Testing-Fuzzing-1d1f189e0047806c8e5cd37781b0a350?pvs=4
11215 : //
11216 : // A fuzzing mode is also supported. In this mode, the test will use a random
11217 : // seed instead of a hardcoded one. Use it in conjunction with `cargo stress`
11218 : // to run multiple instances in parallel:
11219 : //
11220 : // $ RUST_BACKTRACE=1 RUST_LOG=INFO \
11221 : // cargo stress --package=pageserver --features=testing,fuzz-read-path --release -- test_read_path
11222 : #[cfg(feature = "testing")]
11223 : #[tokio::test]
11224 1 : async fn test_read_path() -> anyhow::Result<()> {
11225 : use rand::seq::SliceRandom;
11226 :
11227 1 : let seed = if cfg!(feature = "fuzz-read-path") {
11228 0 : let seed: u64 = thread_rng().r#gen();
11229 0 : seed
11230 : } else {
11231 : // Use a hard-coded seed when not in fuzzing mode.
11232 : // Note that with the current approach results are not reproducible
11233 : // accross platforms and Rust releases.
11234 : const SEED: u64 = 0;
11235 1 : SEED
11236 : };
11237 :
11238 1 : let mut random = StdRng::seed_from_u64(seed);
11239 :
11240 1 : let (queries, will_init_chance, gap_chance) = if cfg!(feature = "fuzz-read-path") {
11241 : const QUERIES: u64 = 5000;
11242 0 : let will_init_chance: u8 = random.gen_range(0..=10);
11243 0 : let gap_chance: u8 = random.gen_range(0..=50);
11244 :
11245 0 : (QUERIES, will_init_chance, gap_chance)
11246 : } else {
11247 : const QUERIES: u64 = 1000;
11248 : const WILL_INIT_CHANCE: u8 = 1;
11249 : const GAP_CHANCE: u8 = 5;
11250 :
11251 1 : (QUERIES, WILL_INIT_CHANCE, GAP_CHANCE)
11252 : };
11253 :
11254 1 : let harness = TenantHarness::create("test_read_path").await?;
11255 1 : let (tenant, ctx) = harness.load().await;
11256 :
11257 1 : tracing::info!("Using random seed: {seed}");
11258 1 : tracing::info!(%will_init_chance, %gap_chance, "Fill params");
11259 :
11260 : // Define the layer map shape. Note that this part is not randomized.
11261 :
11262 : const KEY_DIMENSION_SIZE: u32 = 99;
11263 1 : let start_key = Key::from_hex("110000000033333333444444445500000000").unwrap();
11264 1 : let end_key = start_key.add(KEY_DIMENSION_SIZE);
11265 1 : let total_key_range = start_key..end_key;
11266 1 : let total_key_range_size = end_key.to_i128() - start_key.to_i128();
11267 1 : let total_start_lsn = Lsn(104);
11268 1 : let last_record_lsn = Lsn(504);
11269 :
11270 1 : assert!(total_key_range_size % 3 == 0);
11271 :
11272 1 : let in_memory_layers_shape = vec![
11273 1 : (total_key_range.clone(), Lsn(304)..Lsn(400)),
11274 1 : (total_key_range.clone(), Lsn(400)..last_record_lsn),
11275 : ];
11276 :
11277 1 : let delta_layers_shape = vec![
11278 1 : (
11279 1 : start_key..(start_key.add((total_key_range_size / 3) as u32)),
11280 1 : Lsn(200)..Lsn(304),
11281 1 : ),
11282 1 : (
11283 1 : (start_key.add((total_key_range_size / 3) as u32))
11284 1 : ..(start_key.add((total_key_range_size * 2 / 3) as u32)),
11285 1 : Lsn(200)..Lsn(304),
11286 1 : ),
11287 1 : (
11288 1 : (start_key.add((total_key_range_size * 2 / 3) as u32))
11289 1 : ..(start_key.add(total_key_range_size as u32)),
11290 1 : Lsn(200)..Lsn(304),
11291 1 : ),
11292 : ];
11293 :
11294 1 : let image_layers_shape = vec![
11295 1 : (
11296 1 : start_key.add((total_key_range_size * 2 / 3 - 10) as u32)
11297 1 : ..start_key.add((total_key_range_size * 2 / 3 + 10) as u32),
11298 1 : Lsn(456),
11299 1 : ),
11300 1 : (
11301 1 : start_key.add((total_key_range_size / 3 - 10) as u32)
11302 1 : ..start_key.add((total_key_range_size / 3 + 10) as u32),
11303 1 : Lsn(256),
11304 1 : ),
11305 1 : (total_key_range.clone(), total_start_lsn),
11306 : ];
11307 :
11308 1 : let specification = TestTimelineSpecification {
11309 1 : start_lsn: total_start_lsn,
11310 1 : last_record_lsn,
11311 1 : in_memory_layers_shape,
11312 1 : delta_layers_shape,
11313 1 : image_layers_shape,
11314 1 : gap_chance,
11315 1 : will_init_chance,
11316 1 : };
11317 :
11318 : // Create and randomly fill in the layers according to the specification
11319 1 : let (tline, storage, interesting_lsns) = randomize_timeline(
11320 1 : &tenant,
11321 1 : TIMELINE_ID,
11322 1 : DEFAULT_PG_VERSION,
11323 1 : specification,
11324 1 : &mut random,
11325 1 : &ctx,
11326 1 : )
11327 1 : .await?;
11328 :
11329 : // Now generate queries based on the interesting lsns that we've collected.
11330 : //
11331 : // While there's still room in the query, pick and interesting LSN and a random
11332 : // key. Then roll the dice to see if the next key should also be included in
11333 : // the query. When the roll fails, break the "batch" and pick another point in the
11334 : // (key, LSN) space.
11335 :
11336 : const PICK_NEXT_CHANCE: u8 = 50;
11337 1 : for _ in 0..queries {
11338 1000 : let query = {
11339 1000 : let mut keyspaces_at_lsn: HashMap<Lsn, KeySpaceRandomAccum> = HashMap::default();
11340 1000 : let mut used_keys: HashSet<Key> = HashSet::default();
11341 1 :
11342 22536 : while used_keys.len() < tenant.conf.max_get_vectored_keys.get() {
11343 21536 : let selected_lsn = interesting_lsns.choose(&mut random).expect("not empty");
11344 21536 : let mut selected_key = start_key.add(random.gen_range(0..KEY_DIMENSION_SIZE));
11345 1 :
11346 37614 : while used_keys.len() < tenant.conf.max_get_vectored_keys.get() {
11347 37093 : if used_keys.contains(&selected_key)
11348 32154 : || selected_key >= start_key.add(KEY_DIMENSION_SIZE)
11349 1 : {
11350 5093 : break;
11351 32000 : }
11352 1 :
11353 32000 : keyspaces_at_lsn
11354 32000 : .entry(*selected_lsn)
11355 32000 : .or_default()
11356 32000 : .add_key(selected_key);
11357 32000 : used_keys.insert(selected_key);
11358 1 :
11359 32000 : let pick_next = random.gen_range(0..=100) <= PICK_NEXT_CHANCE;
11360 32000 : if pick_next {
11361 16078 : selected_key = selected_key.next();
11362 16078 : } else {
11363 15922 : break;
11364 1 : }
11365 1 : }
11366 1 : }
11367 1 :
11368 1000 : VersionedKeySpaceQuery::scattered(
11369 1000 : keyspaces_at_lsn
11370 1000 : .into_iter()
11371 11917 : .map(|(lsn, acc)| (lsn, acc.to_keyspace()))
11372 1000 : .collect(),
11373 1 : )
11374 1 : };
11375 1 :
11376 1 : // Run the query and validate the results
11377 1 :
11378 1000 : let results = tline
11379 1000 : .get_vectored(query.clone(), IoConcurrency::Sequential, &ctx)
11380 1000 : .await;
11381 1 :
11382 1000 : let blobs = match results {
11383 1000 : Ok(ok) => ok,
11384 1 : Err(err) => {
11385 1 : panic!("seed={seed} Error returned for query {query}: {err}");
11386 1 : }
11387 1 : };
11388 1 :
11389 32000 : for (key, key_res) in blobs.into_iter() {
11390 32000 : match key_res {
11391 32000 : Ok(blob) => {
11392 32000 : let requested_at_lsn = query.map_key_to_lsn(&key);
11393 32000 : let expected = storage.get(key, requested_at_lsn);
11394 1 :
11395 32000 : if blob != expected {
11396 1 : tracing::error!(
11397 1 : "seed={seed} Mismatch for {key}@{requested_at_lsn} from query: {query}"
11398 1 : );
11399 32000 : }
11400 1 :
11401 32000 : assert_eq!(blob, expected);
11402 1 : }
11403 1 : Err(err) => {
11404 1 : let requested_at_lsn = query.map_key_to_lsn(&key);
11405 1 :
11406 1 : panic!(
11407 1 : "seed={seed} Error returned for {key}@{requested_at_lsn} from query {query}: {err}"
11408 1 : );
11409 1 : }
11410 1 : }
11411 1 : }
11412 1 : }
11413 1 :
11414 1 : Ok(())
11415 1 : }
11416 :
11417 107 : fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
11418 107 : (
11419 107 : k1.is_delta,
11420 107 : k1.key_range.start,
11421 107 : k1.key_range.end,
11422 107 : k1.lsn_range.start,
11423 107 : k1.lsn_range.end,
11424 107 : )
11425 107 : .cmp(&(
11426 107 : k2.is_delta,
11427 107 : k2.key_range.start,
11428 107 : k2.key_range.end,
11429 107 : k2.lsn_range.start,
11430 107 : k2.lsn_range.end,
11431 107 : ))
11432 107 : }
11433 :
11434 12 : async fn inspect_and_sort(
11435 12 : tline: &Arc<Timeline>,
11436 12 : filter: Option<std::ops::Range<Key>>,
11437 12 : ) -> Vec<PersistentLayerKey> {
11438 12 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
11439 12 : if let Some(filter) = filter {
11440 54 : all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
11441 1 : }
11442 12 : all_layers.sort_by(sort_layer_key);
11443 12 : all_layers
11444 12 : }
11445 :
11446 : #[cfg(feature = "testing")]
11447 11 : fn check_layer_map_key_eq(
11448 11 : mut left: Vec<PersistentLayerKey>,
11449 11 : mut right: Vec<PersistentLayerKey>,
11450 11 : ) {
11451 11 : left.sort_by(sort_layer_key);
11452 11 : right.sort_by(sort_layer_key);
11453 11 : if left != right {
11454 0 : eprintln!("---LEFT---");
11455 0 : for left in left.iter() {
11456 0 : eprintln!("{left}");
11457 0 : }
11458 0 : eprintln!("---RIGHT---");
11459 0 : for right in right.iter() {
11460 0 : eprintln!("{right}");
11461 0 : }
11462 0 : assert_eq!(left, right);
11463 11 : }
11464 11 : }
11465 :
11466 : #[cfg(feature = "testing")]
11467 : #[tokio::test]
11468 1 : async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
11469 1 : let harness = TenantHarness::create("test_simple_partial_bottom_most_compaction").await?;
11470 1 : let (tenant, ctx) = harness.load().await;
11471 :
11472 91 : fn get_key(id: u32) -> Key {
11473 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11474 91 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11475 91 : key.field6 = id;
11476 91 : key
11477 91 : }
11478 :
11479 : // img layer at 0x10
11480 1 : let img_layer = (0..10)
11481 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11482 1 : .collect_vec();
11483 :
11484 1 : let delta1 = vec![
11485 1 : (
11486 1 : get_key(1),
11487 1 : Lsn(0x20),
11488 1 : Value::Image(Bytes::from("value 1@0x20")),
11489 1 : ),
11490 1 : (
11491 1 : get_key(2),
11492 1 : Lsn(0x30),
11493 1 : Value::Image(Bytes::from("value 2@0x30")),
11494 1 : ),
11495 1 : (
11496 1 : get_key(3),
11497 1 : Lsn(0x40),
11498 1 : Value::Image(Bytes::from("value 3@0x40")),
11499 1 : ),
11500 : ];
11501 1 : let delta2 = vec![
11502 1 : (
11503 1 : get_key(5),
11504 1 : Lsn(0x20),
11505 1 : Value::Image(Bytes::from("value 5@0x20")),
11506 1 : ),
11507 1 : (
11508 1 : get_key(6),
11509 1 : Lsn(0x20),
11510 1 : Value::Image(Bytes::from("value 6@0x20")),
11511 1 : ),
11512 : ];
11513 1 : let delta3 = vec![
11514 1 : (
11515 1 : get_key(8),
11516 1 : Lsn(0x48),
11517 1 : Value::Image(Bytes::from("value 8@0x48")),
11518 1 : ),
11519 1 : (
11520 1 : get_key(9),
11521 1 : Lsn(0x48),
11522 1 : Value::Image(Bytes::from("value 9@0x48")),
11523 1 : ),
11524 : ];
11525 :
11526 1 : let tline = tenant
11527 1 : .create_test_timeline_with_layers(
11528 1 : TIMELINE_ID,
11529 1 : Lsn(0x10),
11530 1 : DEFAULT_PG_VERSION,
11531 1 : &ctx,
11532 1 : vec![], // in-memory layers
11533 1 : vec![
11534 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
11535 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
11536 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
11537 1 : ], // delta layers
11538 1 : vec![(Lsn(0x10), img_layer)], // image layers
11539 1 : Lsn(0x50),
11540 1 : )
11541 1 : .await?;
11542 :
11543 : {
11544 1 : tline
11545 1 : .applied_gc_cutoff_lsn
11546 1 : .lock_for_write()
11547 1 : .store_and_unlock(Lsn(0x30))
11548 1 : .wait()
11549 1 : .await;
11550 : // Update GC info
11551 1 : let mut guard = tline.gc_info.write().unwrap();
11552 1 : *guard = GcInfo {
11553 1 : retain_lsns: vec![(Lsn(0x20), tline.timeline_id, MaybeOffloaded::No)],
11554 1 : cutoffs: GcCutoffs {
11555 1 : time: Some(Lsn(0x30)),
11556 1 : space: Lsn(0x30),
11557 1 : },
11558 1 : leases: Default::default(),
11559 1 : within_ancestor_pitr: false,
11560 1 : };
11561 : }
11562 :
11563 1 : let cancel = CancellationToken::new();
11564 :
11565 : // Do a partial compaction on key range 0..2
11566 1 : tline
11567 1 : .compact_with_gc(
11568 1 : &cancel,
11569 1 : CompactOptions {
11570 1 : flags: EnumSet::new(),
11571 1 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
11572 1 : ..Default::default()
11573 1 : },
11574 1 : &ctx,
11575 1 : )
11576 1 : .await
11577 1 : .unwrap();
11578 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11579 1 : check_layer_map_key_eq(
11580 1 : all_layers,
11581 1 : vec![
11582 : // newly-generated image layer for the partial compaction range 0-2
11583 1 : PersistentLayerKey {
11584 1 : key_range: get_key(0)..get_key(2),
11585 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11586 1 : is_delta: false,
11587 1 : },
11588 1 : PersistentLayerKey {
11589 1 : key_range: get_key(0)..get_key(10),
11590 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11591 1 : is_delta: false,
11592 1 : },
11593 : // delta1 is split and the second part is rewritten
11594 1 : PersistentLayerKey {
11595 1 : key_range: get_key(2)..get_key(4),
11596 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11597 1 : is_delta: true,
11598 1 : },
11599 1 : PersistentLayerKey {
11600 1 : key_range: get_key(5)..get_key(7),
11601 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11602 1 : is_delta: true,
11603 1 : },
11604 1 : PersistentLayerKey {
11605 1 : key_range: get_key(8)..get_key(10),
11606 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11607 1 : is_delta: true,
11608 1 : },
11609 : ],
11610 : );
11611 :
11612 : // Do a partial compaction on key range 2..4
11613 1 : tline
11614 1 : .compact_with_gc(
11615 1 : &cancel,
11616 1 : CompactOptions {
11617 1 : flags: EnumSet::new(),
11618 1 : compact_key_range: Some((get_key(2)..get_key(4)).into()),
11619 1 : ..Default::default()
11620 1 : },
11621 1 : &ctx,
11622 1 : )
11623 1 : .await
11624 1 : .unwrap();
11625 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11626 1 : check_layer_map_key_eq(
11627 1 : all_layers,
11628 1 : vec![
11629 1 : PersistentLayerKey {
11630 1 : key_range: get_key(0)..get_key(2),
11631 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11632 1 : is_delta: false,
11633 1 : },
11634 1 : PersistentLayerKey {
11635 1 : key_range: get_key(0)..get_key(10),
11636 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11637 1 : is_delta: false,
11638 1 : },
11639 : // image layer generated for the compaction range 2-4
11640 1 : PersistentLayerKey {
11641 1 : key_range: get_key(2)..get_key(4),
11642 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11643 1 : is_delta: false,
11644 1 : },
11645 : // we have key2/key3 above the retain_lsn, so we still need this delta layer
11646 1 : PersistentLayerKey {
11647 1 : key_range: get_key(2)..get_key(4),
11648 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11649 1 : is_delta: true,
11650 1 : },
11651 1 : PersistentLayerKey {
11652 1 : key_range: get_key(5)..get_key(7),
11653 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11654 1 : is_delta: true,
11655 1 : },
11656 1 : PersistentLayerKey {
11657 1 : key_range: get_key(8)..get_key(10),
11658 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11659 1 : is_delta: true,
11660 1 : },
11661 : ],
11662 : );
11663 :
11664 : // Do a partial compaction on key range 4..9
11665 1 : tline
11666 1 : .compact_with_gc(
11667 1 : &cancel,
11668 1 : CompactOptions {
11669 1 : flags: EnumSet::new(),
11670 1 : compact_key_range: Some((get_key(4)..get_key(9)).into()),
11671 1 : ..Default::default()
11672 1 : },
11673 1 : &ctx,
11674 1 : )
11675 1 : .await
11676 1 : .unwrap();
11677 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11678 1 : check_layer_map_key_eq(
11679 1 : all_layers,
11680 1 : vec![
11681 1 : PersistentLayerKey {
11682 1 : key_range: get_key(0)..get_key(2),
11683 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11684 1 : is_delta: false,
11685 1 : },
11686 1 : PersistentLayerKey {
11687 1 : key_range: get_key(0)..get_key(10),
11688 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11689 1 : is_delta: false,
11690 1 : },
11691 1 : PersistentLayerKey {
11692 1 : key_range: get_key(2)..get_key(4),
11693 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11694 1 : is_delta: false,
11695 1 : },
11696 1 : PersistentLayerKey {
11697 1 : key_range: get_key(2)..get_key(4),
11698 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11699 1 : is_delta: true,
11700 1 : },
11701 : // image layer generated for this compaction range
11702 1 : PersistentLayerKey {
11703 1 : key_range: get_key(4)..get_key(9),
11704 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11705 1 : is_delta: false,
11706 1 : },
11707 1 : PersistentLayerKey {
11708 1 : key_range: get_key(8)..get_key(10),
11709 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11710 1 : is_delta: true,
11711 1 : },
11712 : ],
11713 : );
11714 :
11715 : // Do a partial compaction on key range 9..10
11716 1 : tline
11717 1 : .compact_with_gc(
11718 1 : &cancel,
11719 1 : CompactOptions {
11720 1 : flags: EnumSet::new(),
11721 1 : compact_key_range: Some((get_key(9)..get_key(10)).into()),
11722 1 : ..Default::default()
11723 1 : },
11724 1 : &ctx,
11725 1 : )
11726 1 : .await
11727 1 : .unwrap();
11728 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11729 1 : check_layer_map_key_eq(
11730 1 : all_layers,
11731 1 : vec![
11732 1 : PersistentLayerKey {
11733 1 : key_range: get_key(0)..get_key(2),
11734 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11735 1 : is_delta: false,
11736 1 : },
11737 1 : PersistentLayerKey {
11738 1 : key_range: get_key(0)..get_key(10),
11739 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
11740 1 : is_delta: false,
11741 1 : },
11742 1 : PersistentLayerKey {
11743 1 : key_range: get_key(2)..get_key(4),
11744 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11745 1 : is_delta: false,
11746 1 : },
11747 1 : PersistentLayerKey {
11748 1 : key_range: get_key(2)..get_key(4),
11749 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11750 1 : is_delta: true,
11751 1 : },
11752 1 : PersistentLayerKey {
11753 1 : key_range: get_key(4)..get_key(9),
11754 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11755 1 : is_delta: false,
11756 1 : },
11757 : // image layer generated for the compaction range
11758 1 : PersistentLayerKey {
11759 1 : key_range: get_key(9)..get_key(10),
11760 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11761 1 : is_delta: false,
11762 1 : },
11763 1 : PersistentLayerKey {
11764 1 : key_range: get_key(8)..get_key(10),
11765 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11766 1 : is_delta: true,
11767 1 : },
11768 : ],
11769 : );
11770 :
11771 : // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
11772 1 : tline
11773 1 : .compact_with_gc(
11774 1 : &cancel,
11775 1 : CompactOptions {
11776 1 : flags: EnumSet::new(),
11777 1 : compact_key_range: Some((get_key(0)..get_key(10)).into()),
11778 1 : ..Default::default()
11779 1 : },
11780 1 : &ctx,
11781 1 : )
11782 1 : .await
11783 1 : .unwrap();
11784 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
11785 1 : check_layer_map_key_eq(
11786 1 : all_layers,
11787 1 : vec![
11788 : // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
11789 1 : PersistentLayerKey {
11790 1 : key_range: get_key(0)..get_key(10),
11791 1 : lsn_range: Lsn(0x20)..Lsn(0x21),
11792 1 : is_delta: false,
11793 1 : },
11794 1 : PersistentLayerKey {
11795 1 : key_range: get_key(2)..get_key(4),
11796 1 : lsn_range: Lsn(0x20)..Lsn(0x48),
11797 1 : is_delta: true,
11798 1 : },
11799 1 : PersistentLayerKey {
11800 1 : key_range: get_key(8)..get_key(10),
11801 1 : lsn_range: Lsn(0x48)..Lsn(0x50),
11802 1 : is_delta: true,
11803 1 : },
11804 : ],
11805 : );
11806 2 : Ok(())
11807 1 : }
11808 :
11809 : #[cfg(feature = "testing")]
11810 : #[tokio::test]
11811 1 : async fn test_timeline_offload_retain_lsn() -> anyhow::Result<()> {
11812 1 : let harness = TenantHarness::create("test_timeline_offload_retain_lsn")
11813 1 : .await
11814 1 : .unwrap();
11815 1 : let (tenant, ctx) = harness.load().await;
11816 1 : let tline_parent = tenant
11817 1 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
11818 1 : .await
11819 1 : .unwrap();
11820 1 : let tline_child = tenant
11821 1 : .branch_timeline_test(&tline_parent, NEW_TIMELINE_ID, Some(Lsn(0x20)), &ctx)
11822 1 : .await
11823 1 : .unwrap();
11824 : {
11825 1 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11826 1 : assert_eq!(
11827 1 : gc_info_parent.retain_lsns,
11828 1 : vec![(Lsn(0x20), tline_child.timeline_id, MaybeOffloaded::No)]
11829 : );
11830 : }
11831 : // We have to directly call the remote_client instead of using the archive function to avoid constructing broker client...
11832 1 : tline_child
11833 1 : .remote_client
11834 1 : .schedule_index_upload_for_timeline_archival_state(TimelineArchivalState::Archived)
11835 1 : .unwrap();
11836 1 : tline_child.remote_client.wait_completion().await.unwrap();
11837 1 : offload_timeline(&tenant, &tline_child)
11838 1 : .instrument(tracing::info_span!(parent: None, "offload_test", tenant_id=%"test", shard_id=%"test", timeline_id=%"test"))
11839 1 : .await.unwrap();
11840 1 : let child_timeline_id = tline_child.timeline_id;
11841 1 : Arc::try_unwrap(tline_child).unwrap();
11842 :
11843 : {
11844 1 : let gc_info_parent = tline_parent.gc_info.read().unwrap();
11845 1 : assert_eq!(
11846 1 : gc_info_parent.retain_lsns,
11847 1 : vec![(Lsn(0x20), child_timeline_id, MaybeOffloaded::Yes)]
11848 : );
11849 : }
11850 :
11851 1 : tenant
11852 1 : .get_offloaded_timeline(child_timeline_id)
11853 1 : .unwrap()
11854 1 : .defuse_for_tenant_drop();
11855 :
11856 2 : Ok(())
11857 1 : }
11858 :
11859 : #[cfg(feature = "testing")]
11860 : #[tokio::test]
11861 1 : async fn test_simple_bottom_most_compaction_above_lsn() -> anyhow::Result<()> {
11862 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_above_lsn").await?;
11863 1 : let (tenant, ctx) = harness.load().await;
11864 :
11865 148 : fn get_key(id: u32) -> Key {
11866 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
11867 148 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
11868 148 : key.field6 = id;
11869 148 : key
11870 148 : }
11871 :
11872 1 : let img_layer = (0..10)
11873 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
11874 1 : .collect_vec();
11875 :
11876 1 : let delta1 = vec![(
11877 1 : get_key(1),
11878 1 : Lsn(0x20),
11879 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
11880 1 : )];
11881 1 : let delta4 = vec![(
11882 1 : get_key(1),
11883 1 : Lsn(0x28),
11884 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
11885 1 : )];
11886 1 : let delta2 = vec![
11887 1 : (
11888 1 : get_key(1),
11889 1 : Lsn(0x30),
11890 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
11891 1 : ),
11892 1 : (
11893 1 : get_key(1),
11894 1 : Lsn(0x38),
11895 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
11896 1 : ),
11897 : ];
11898 1 : let delta3 = vec![
11899 1 : (
11900 1 : get_key(8),
11901 1 : Lsn(0x48),
11902 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11903 1 : ),
11904 1 : (
11905 1 : get_key(9),
11906 1 : Lsn(0x48),
11907 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
11908 1 : ),
11909 : ];
11910 :
11911 1 : let tline = tenant
11912 1 : .create_test_timeline_with_layers(
11913 1 : TIMELINE_ID,
11914 1 : Lsn(0x10),
11915 1 : DEFAULT_PG_VERSION,
11916 1 : &ctx,
11917 1 : vec![], // in-memory layers
11918 1 : vec![
11919 1 : // delta1/2/4 only contain a single key but multiple updates
11920 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
11921 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
11922 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
11923 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
11924 1 : ], // delta layers
11925 1 : vec![(Lsn(0x10), img_layer)], // image layers
11926 1 : Lsn(0x50),
11927 1 : )
11928 1 : .await?;
11929 : {
11930 1 : tline
11931 1 : .applied_gc_cutoff_lsn
11932 1 : .lock_for_write()
11933 1 : .store_and_unlock(Lsn(0x30))
11934 1 : .wait()
11935 1 : .await;
11936 : // Update GC info
11937 1 : let mut guard = tline.gc_info.write().unwrap();
11938 1 : *guard = GcInfo {
11939 1 : retain_lsns: vec![
11940 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
11941 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
11942 1 : ],
11943 1 : cutoffs: GcCutoffs {
11944 1 : time: Some(Lsn(0x30)),
11945 1 : space: Lsn(0x30),
11946 1 : },
11947 1 : leases: Default::default(),
11948 1 : within_ancestor_pitr: false,
11949 1 : };
11950 : }
11951 :
11952 1 : let expected_result = [
11953 1 : Bytes::from_static(b"value 0@0x10"),
11954 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
11955 1 : Bytes::from_static(b"value 2@0x10"),
11956 1 : Bytes::from_static(b"value 3@0x10"),
11957 1 : Bytes::from_static(b"value 4@0x10"),
11958 1 : Bytes::from_static(b"value 5@0x10"),
11959 1 : Bytes::from_static(b"value 6@0x10"),
11960 1 : Bytes::from_static(b"value 7@0x10"),
11961 1 : Bytes::from_static(b"value 8@0x10@0x48"),
11962 1 : Bytes::from_static(b"value 9@0x10@0x48"),
11963 1 : ];
11964 :
11965 1 : let expected_result_at_gc_horizon = [
11966 1 : Bytes::from_static(b"value 0@0x10"),
11967 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
11968 1 : Bytes::from_static(b"value 2@0x10"),
11969 1 : Bytes::from_static(b"value 3@0x10"),
11970 1 : Bytes::from_static(b"value 4@0x10"),
11971 1 : Bytes::from_static(b"value 5@0x10"),
11972 1 : Bytes::from_static(b"value 6@0x10"),
11973 1 : Bytes::from_static(b"value 7@0x10"),
11974 1 : Bytes::from_static(b"value 8@0x10"),
11975 1 : Bytes::from_static(b"value 9@0x10"),
11976 1 : ];
11977 :
11978 1 : let expected_result_at_lsn_20 = [
11979 1 : Bytes::from_static(b"value 0@0x10"),
11980 1 : Bytes::from_static(b"value 1@0x10@0x20"),
11981 1 : Bytes::from_static(b"value 2@0x10"),
11982 1 : Bytes::from_static(b"value 3@0x10"),
11983 1 : Bytes::from_static(b"value 4@0x10"),
11984 1 : Bytes::from_static(b"value 5@0x10"),
11985 1 : Bytes::from_static(b"value 6@0x10"),
11986 1 : Bytes::from_static(b"value 7@0x10"),
11987 1 : Bytes::from_static(b"value 8@0x10"),
11988 1 : Bytes::from_static(b"value 9@0x10"),
11989 1 : ];
11990 :
11991 1 : let expected_result_at_lsn_10 = [
11992 1 : Bytes::from_static(b"value 0@0x10"),
11993 1 : Bytes::from_static(b"value 1@0x10"),
11994 1 : Bytes::from_static(b"value 2@0x10"),
11995 1 : Bytes::from_static(b"value 3@0x10"),
11996 1 : Bytes::from_static(b"value 4@0x10"),
11997 1 : Bytes::from_static(b"value 5@0x10"),
11998 1 : Bytes::from_static(b"value 6@0x10"),
11999 1 : Bytes::from_static(b"value 7@0x10"),
12000 1 : Bytes::from_static(b"value 8@0x10"),
12001 1 : Bytes::from_static(b"value 9@0x10"),
12002 1 : ];
12003 :
12004 3 : let verify_result = || async {
12005 3 : let gc_horizon = {
12006 3 : let gc_info = tline.gc_info.read().unwrap();
12007 3 : gc_info.cutoffs.time.unwrap_or_default()
12008 : };
12009 33 : for idx in 0..10 {
12010 30 : assert_eq!(
12011 30 : tline
12012 30 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
12013 30 : .await
12014 30 : .unwrap(),
12015 30 : &expected_result[idx]
12016 : );
12017 30 : assert_eq!(
12018 30 : tline
12019 30 : .get(get_key(idx as u32), gc_horizon, &ctx)
12020 30 : .await
12021 30 : .unwrap(),
12022 30 : &expected_result_at_gc_horizon[idx]
12023 : );
12024 30 : assert_eq!(
12025 30 : tline
12026 30 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
12027 30 : .await
12028 30 : .unwrap(),
12029 30 : &expected_result_at_lsn_20[idx]
12030 : );
12031 30 : assert_eq!(
12032 30 : tline
12033 30 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
12034 30 : .await
12035 30 : .unwrap(),
12036 30 : &expected_result_at_lsn_10[idx]
12037 : );
12038 : }
12039 6 : };
12040 :
12041 1 : verify_result().await;
12042 :
12043 1 : let cancel = CancellationToken::new();
12044 1 : tline
12045 1 : .compact_with_gc(
12046 1 : &cancel,
12047 1 : CompactOptions {
12048 1 : compact_lsn_range: Some(CompactLsnRange::above(Lsn(0x28))),
12049 1 : ..Default::default()
12050 1 : },
12051 1 : &ctx,
12052 1 : )
12053 1 : .await
12054 1 : .unwrap();
12055 1 : verify_result().await;
12056 :
12057 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12058 1 : check_layer_map_key_eq(
12059 1 : all_layers,
12060 1 : vec![
12061 : // The original image layer, not compacted
12062 1 : PersistentLayerKey {
12063 1 : key_range: get_key(0)..get_key(10),
12064 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12065 1 : is_delta: false,
12066 1 : },
12067 : // Delta layer below the specified above_lsn not compacted
12068 1 : PersistentLayerKey {
12069 1 : key_range: get_key(1)..get_key(2),
12070 1 : lsn_range: Lsn(0x20)..Lsn(0x28),
12071 1 : is_delta: true,
12072 1 : },
12073 : // Delta layer compacted above the LSN
12074 1 : PersistentLayerKey {
12075 1 : key_range: get_key(1)..get_key(10),
12076 1 : lsn_range: Lsn(0x28)..Lsn(0x50),
12077 1 : is_delta: true,
12078 1 : },
12079 : ],
12080 : );
12081 :
12082 : // compact again
12083 1 : tline
12084 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
12085 1 : .await
12086 1 : .unwrap();
12087 1 : verify_result().await;
12088 :
12089 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12090 1 : check_layer_map_key_eq(
12091 1 : all_layers,
12092 1 : vec![
12093 : // The compacted image layer (full key range)
12094 1 : PersistentLayerKey {
12095 1 : key_range: Key::MIN..Key::MAX,
12096 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12097 1 : is_delta: false,
12098 1 : },
12099 : // All other data in the delta layer
12100 1 : PersistentLayerKey {
12101 1 : key_range: get_key(1)..get_key(10),
12102 1 : lsn_range: Lsn(0x10)..Lsn(0x50),
12103 1 : is_delta: true,
12104 1 : },
12105 : ],
12106 : );
12107 :
12108 2 : Ok(())
12109 1 : }
12110 :
12111 : #[cfg(feature = "testing")]
12112 : #[tokio::test]
12113 1 : async fn test_simple_bottom_most_compaction_rectangle() -> anyhow::Result<()> {
12114 1 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_rectangle").await?;
12115 1 : let (tenant, ctx) = harness.load().await;
12116 :
12117 254 : fn get_key(id: u32) -> Key {
12118 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
12119 254 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
12120 254 : key.field6 = id;
12121 254 : key
12122 254 : }
12123 :
12124 1 : let img_layer = (0..10)
12125 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
12126 1 : .collect_vec();
12127 :
12128 1 : let delta1 = vec![(
12129 1 : get_key(1),
12130 1 : Lsn(0x20),
12131 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
12132 1 : )];
12133 1 : let delta4 = vec![(
12134 1 : get_key(1),
12135 1 : Lsn(0x28),
12136 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
12137 1 : )];
12138 1 : let delta2 = vec![
12139 1 : (
12140 1 : get_key(1),
12141 1 : Lsn(0x30),
12142 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
12143 1 : ),
12144 1 : (
12145 1 : get_key(1),
12146 1 : Lsn(0x38),
12147 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x38")),
12148 1 : ),
12149 : ];
12150 1 : let delta3 = vec![
12151 1 : (
12152 1 : get_key(8),
12153 1 : Lsn(0x48),
12154 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
12155 1 : ),
12156 1 : (
12157 1 : get_key(9),
12158 1 : Lsn(0x48),
12159 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
12160 1 : ),
12161 : ];
12162 :
12163 1 : let tline = tenant
12164 1 : .create_test_timeline_with_layers(
12165 1 : TIMELINE_ID,
12166 1 : Lsn(0x10),
12167 1 : DEFAULT_PG_VERSION,
12168 1 : &ctx,
12169 1 : vec![], // in-memory layers
12170 1 : vec![
12171 1 : // delta1/2/4 only contain a single key but multiple updates
12172 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x28), delta1),
12173 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta2),
12174 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x28)..Lsn(0x30), delta4),
12175 1 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x30)..Lsn(0x50), delta3),
12176 1 : ], // delta layers
12177 1 : vec![(Lsn(0x10), img_layer)], // image layers
12178 1 : Lsn(0x50),
12179 1 : )
12180 1 : .await?;
12181 : {
12182 1 : tline
12183 1 : .applied_gc_cutoff_lsn
12184 1 : .lock_for_write()
12185 1 : .store_and_unlock(Lsn(0x30))
12186 1 : .wait()
12187 1 : .await;
12188 : // Update GC info
12189 1 : let mut guard = tline.gc_info.write().unwrap();
12190 1 : *guard = GcInfo {
12191 1 : retain_lsns: vec![
12192 1 : (Lsn(0x10), tline.timeline_id, MaybeOffloaded::No),
12193 1 : (Lsn(0x20), tline.timeline_id, MaybeOffloaded::No),
12194 1 : ],
12195 1 : cutoffs: GcCutoffs {
12196 1 : time: Some(Lsn(0x30)),
12197 1 : space: Lsn(0x30),
12198 1 : },
12199 1 : leases: Default::default(),
12200 1 : within_ancestor_pitr: false,
12201 1 : };
12202 : }
12203 :
12204 1 : let expected_result = [
12205 1 : Bytes::from_static(b"value 0@0x10"),
12206 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30@0x38"),
12207 1 : Bytes::from_static(b"value 2@0x10"),
12208 1 : Bytes::from_static(b"value 3@0x10"),
12209 1 : Bytes::from_static(b"value 4@0x10"),
12210 1 : Bytes::from_static(b"value 5@0x10"),
12211 1 : Bytes::from_static(b"value 6@0x10"),
12212 1 : Bytes::from_static(b"value 7@0x10"),
12213 1 : Bytes::from_static(b"value 8@0x10@0x48"),
12214 1 : Bytes::from_static(b"value 9@0x10@0x48"),
12215 1 : ];
12216 :
12217 1 : let expected_result_at_gc_horizon = [
12218 1 : Bytes::from_static(b"value 0@0x10"),
12219 1 : Bytes::from_static(b"value 1@0x10@0x20@0x28@0x30"),
12220 1 : Bytes::from_static(b"value 2@0x10"),
12221 1 : Bytes::from_static(b"value 3@0x10"),
12222 1 : Bytes::from_static(b"value 4@0x10"),
12223 1 : Bytes::from_static(b"value 5@0x10"),
12224 1 : Bytes::from_static(b"value 6@0x10"),
12225 1 : Bytes::from_static(b"value 7@0x10"),
12226 1 : Bytes::from_static(b"value 8@0x10"),
12227 1 : Bytes::from_static(b"value 9@0x10"),
12228 1 : ];
12229 :
12230 1 : let expected_result_at_lsn_20 = [
12231 1 : Bytes::from_static(b"value 0@0x10"),
12232 1 : Bytes::from_static(b"value 1@0x10@0x20"),
12233 1 : Bytes::from_static(b"value 2@0x10"),
12234 1 : Bytes::from_static(b"value 3@0x10"),
12235 1 : Bytes::from_static(b"value 4@0x10"),
12236 1 : Bytes::from_static(b"value 5@0x10"),
12237 1 : Bytes::from_static(b"value 6@0x10"),
12238 1 : Bytes::from_static(b"value 7@0x10"),
12239 1 : Bytes::from_static(b"value 8@0x10"),
12240 1 : Bytes::from_static(b"value 9@0x10"),
12241 1 : ];
12242 :
12243 1 : let expected_result_at_lsn_10 = [
12244 1 : Bytes::from_static(b"value 0@0x10"),
12245 1 : Bytes::from_static(b"value 1@0x10"),
12246 1 : Bytes::from_static(b"value 2@0x10"),
12247 1 : Bytes::from_static(b"value 3@0x10"),
12248 1 : Bytes::from_static(b"value 4@0x10"),
12249 1 : Bytes::from_static(b"value 5@0x10"),
12250 1 : Bytes::from_static(b"value 6@0x10"),
12251 1 : Bytes::from_static(b"value 7@0x10"),
12252 1 : Bytes::from_static(b"value 8@0x10"),
12253 1 : Bytes::from_static(b"value 9@0x10"),
12254 1 : ];
12255 :
12256 5 : let verify_result = || async {
12257 5 : let gc_horizon = {
12258 5 : let gc_info = tline.gc_info.read().unwrap();
12259 5 : gc_info.cutoffs.time.unwrap_or_default()
12260 : };
12261 55 : for idx in 0..10 {
12262 50 : assert_eq!(
12263 50 : tline
12264 50 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
12265 50 : .await
12266 50 : .unwrap(),
12267 50 : &expected_result[idx]
12268 : );
12269 50 : assert_eq!(
12270 50 : tline
12271 50 : .get(get_key(idx as u32), gc_horizon, &ctx)
12272 50 : .await
12273 50 : .unwrap(),
12274 50 : &expected_result_at_gc_horizon[idx]
12275 : );
12276 50 : assert_eq!(
12277 50 : tline
12278 50 : .get(get_key(idx as u32), Lsn(0x20), &ctx)
12279 50 : .await
12280 50 : .unwrap(),
12281 50 : &expected_result_at_lsn_20[idx]
12282 : );
12283 50 : assert_eq!(
12284 50 : tline
12285 50 : .get(get_key(idx as u32), Lsn(0x10), &ctx)
12286 50 : .await
12287 50 : .unwrap(),
12288 50 : &expected_result_at_lsn_10[idx]
12289 : );
12290 : }
12291 10 : };
12292 :
12293 1 : verify_result().await;
12294 :
12295 1 : let cancel = CancellationToken::new();
12296 :
12297 1 : tline
12298 1 : .compact_with_gc(
12299 1 : &cancel,
12300 1 : CompactOptions {
12301 1 : compact_key_range: Some((get_key(0)..get_key(2)).into()),
12302 1 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x28)).into()),
12303 1 : ..Default::default()
12304 1 : },
12305 1 : &ctx,
12306 1 : )
12307 1 : .await
12308 1 : .unwrap();
12309 1 : verify_result().await;
12310 :
12311 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12312 1 : check_layer_map_key_eq(
12313 1 : all_layers,
12314 1 : vec![
12315 : // The original image layer, not compacted
12316 1 : PersistentLayerKey {
12317 1 : key_range: get_key(0)..get_key(10),
12318 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12319 1 : is_delta: false,
12320 1 : },
12321 : // According the selection logic, we select all layers with start key <= 0x28, so we would merge the layer 0x20-0x28 and
12322 : // the layer 0x28-0x30 into one.
12323 1 : PersistentLayerKey {
12324 1 : key_range: get_key(1)..get_key(2),
12325 1 : lsn_range: Lsn(0x20)..Lsn(0x30),
12326 1 : is_delta: true,
12327 1 : },
12328 : // Above the upper bound and untouched
12329 1 : PersistentLayerKey {
12330 1 : key_range: get_key(1)..get_key(2),
12331 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12332 1 : is_delta: true,
12333 1 : },
12334 : // This layer is untouched
12335 1 : PersistentLayerKey {
12336 1 : key_range: get_key(8)..get_key(10),
12337 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12338 1 : is_delta: true,
12339 1 : },
12340 : ],
12341 : );
12342 :
12343 1 : tline
12344 1 : .compact_with_gc(
12345 1 : &cancel,
12346 1 : CompactOptions {
12347 1 : compact_key_range: Some((get_key(3)..get_key(8)).into()),
12348 1 : compact_lsn_range: Some((Lsn(0x28)..Lsn(0x40)).into()),
12349 1 : ..Default::default()
12350 1 : },
12351 1 : &ctx,
12352 1 : )
12353 1 : .await
12354 1 : .unwrap();
12355 1 : verify_result().await;
12356 :
12357 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12358 1 : check_layer_map_key_eq(
12359 1 : all_layers,
12360 1 : vec![
12361 : // The original image layer, not compacted
12362 1 : PersistentLayerKey {
12363 1 : key_range: get_key(0)..get_key(10),
12364 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12365 1 : is_delta: false,
12366 1 : },
12367 : // Not in the compaction key range, uncompacted
12368 1 : PersistentLayerKey {
12369 1 : key_range: get_key(1)..get_key(2),
12370 1 : lsn_range: Lsn(0x20)..Lsn(0x30),
12371 1 : is_delta: true,
12372 1 : },
12373 : // Not in the compaction key range, uncompacted but need rewrite because the delta layer overlaps with the range
12374 1 : PersistentLayerKey {
12375 1 : key_range: get_key(1)..get_key(2),
12376 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12377 1 : is_delta: true,
12378 1 : },
12379 : // Note that when we specify the LSN upper bound to be 0x40, the compaction algorithm will not try to cut the layer
12380 : // horizontally in half. Instead, it will include all LSNs that overlap with 0x40. So the real max_lsn of the compaction
12381 : // becomes 0x50.
12382 1 : PersistentLayerKey {
12383 1 : key_range: get_key(8)..get_key(10),
12384 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12385 1 : is_delta: true,
12386 1 : },
12387 : ],
12388 : );
12389 :
12390 : // compact again
12391 1 : tline
12392 1 : .compact_with_gc(
12393 1 : &cancel,
12394 1 : CompactOptions {
12395 1 : compact_key_range: Some((get_key(0)..get_key(5)).into()),
12396 1 : compact_lsn_range: Some((Lsn(0x20)..Lsn(0x50)).into()),
12397 1 : ..Default::default()
12398 1 : },
12399 1 : &ctx,
12400 1 : )
12401 1 : .await
12402 1 : .unwrap();
12403 1 : verify_result().await;
12404 :
12405 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12406 1 : check_layer_map_key_eq(
12407 1 : all_layers,
12408 1 : vec![
12409 : // The original image layer, not compacted
12410 1 : PersistentLayerKey {
12411 1 : key_range: get_key(0)..get_key(10),
12412 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12413 1 : is_delta: false,
12414 1 : },
12415 : // The range gets compacted
12416 1 : PersistentLayerKey {
12417 1 : key_range: get_key(1)..get_key(2),
12418 1 : lsn_range: Lsn(0x20)..Lsn(0x50),
12419 1 : is_delta: true,
12420 1 : },
12421 : // Not touched during this iteration of compaction
12422 1 : PersistentLayerKey {
12423 1 : key_range: get_key(8)..get_key(10),
12424 1 : lsn_range: Lsn(0x30)..Lsn(0x50),
12425 1 : is_delta: true,
12426 1 : },
12427 : ],
12428 : );
12429 :
12430 : // final full compaction
12431 1 : tline
12432 1 : .compact_with_gc(&cancel, CompactOptions::default(), &ctx)
12433 1 : .await
12434 1 : .unwrap();
12435 1 : verify_result().await;
12436 :
12437 1 : let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
12438 1 : check_layer_map_key_eq(
12439 1 : all_layers,
12440 1 : vec![
12441 : // The compacted image layer (full key range)
12442 1 : PersistentLayerKey {
12443 1 : key_range: Key::MIN..Key::MAX,
12444 1 : lsn_range: Lsn(0x10)..Lsn(0x11),
12445 1 : is_delta: false,
12446 1 : },
12447 : // All other data in the delta layer
12448 1 : PersistentLayerKey {
12449 1 : key_range: get_key(1)..get_key(10),
12450 1 : lsn_range: Lsn(0x10)..Lsn(0x50),
12451 1 : is_delta: true,
12452 1 : },
12453 : ],
12454 : );
12455 :
12456 2 : Ok(())
12457 1 : }
12458 :
12459 : #[cfg(feature = "testing")]
12460 : #[tokio::test]
12461 1 : async fn test_bottom_most_compation_redo_failure() -> anyhow::Result<()> {
12462 1 : let harness = TenantHarness::create("test_bottom_most_compation_redo_failure").await?;
12463 1 : let (tenant, ctx) = harness.load().await;
12464 :
12465 13 : fn get_key(id: u32) -> Key {
12466 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
12467 13 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
12468 13 : key.field6 = id;
12469 13 : key
12470 13 : }
12471 :
12472 1 : let img_layer = (0..10)
12473 10 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
12474 1 : .collect_vec();
12475 :
12476 1 : let delta1 = vec![
12477 1 : (
12478 1 : get_key(1),
12479 1 : Lsn(0x20),
12480 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
12481 1 : ),
12482 1 : (
12483 1 : get_key(1),
12484 1 : Lsn(0x24),
12485 1 : Value::WalRecord(NeonWalRecord::wal_append("@0x24")),
12486 1 : ),
12487 1 : (
12488 1 : get_key(1),
12489 1 : Lsn(0x28),
12490 1 : // This record will fail to redo
12491 1 : Value::WalRecord(NeonWalRecord::wal_append_conditional("@0x28", "???")),
12492 1 : ),
12493 : ];
12494 :
12495 1 : let tline = tenant
12496 1 : .create_test_timeline_with_layers(
12497 1 : TIMELINE_ID,
12498 1 : Lsn(0x10),
12499 1 : DEFAULT_PG_VERSION,
12500 1 : &ctx,
12501 1 : vec![], // in-memory layers
12502 1 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
12503 1 : Lsn(0x20)..Lsn(0x30),
12504 1 : delta1,
12505 1 : )], // delta layers
12506 1 : vec![(Lsn(0x10), img_layer)], // image layers
12507 1 : Lsn(0x50),
12508 1 : )
12509 1 : .await?;
12510 : {
12511 1 : tline
12512 1 : .applied_gc_cutoff_lsn
12513 1 : .lock_for_write()
12514 1 : .store_and_unlock(Lsn(0x30))
12515 1 : .wait()
12516 1 : .await;
12517 : // Update GC info
12518 1 : let mut guard = tline.gc_info.write().unwrap();
12519 1 : *guard = GcInfo {
12520 1 : retain_lsns: vec![],
12521 1 : cutoffs: GcCutoffs {
12522 1 : time: Some(Lsn(0x30)),
12523 1 : space: Lsn(0x30),
12524 1 : },
12525 1 : leases: Default::default(),
12526 1 : within_ancestor_pitr: false,
12527 1 : };
12528 : }
12529 :
12530 1 : let cancel = CancellationToken::new();
12531 :
12532 : // Compaction will fail, but should not fire any critical error.
12533 : // Gc-compaction currently cannot figure out what keys are not in the keyspace during the compaction
12534 : // process. It will always try to redo the logs it reads and if it doesn't work, fail the entire
12535 : // compaction job. Tracked in <https://github.com/neondatabase/neon/issues/10395>.
12536 1 : let res = tline
12537 1 : .compact_with_gc(
12538 1 : &cancel,
12539 1 : CompactOptions {
12540 1 : compact_key_range: None,
12541 1 : compact_lsn_range: None,
12542 1 : ..Default::default()
12543 1 : },
12544 1 : &ctx,
12545 1 : )
12546 1 : .await;
12547 1 : assert!(res.is_err());
12548 :
12549 2 : Ok(())
12550 1 : }
12551 :
12552 : #[cfg(feature = "testing")]
12553 : #[tokio::test]
12554 1 : async fn test_synthetic_size_calculation_with_invisible_branches() -> anyhow::Result<()> {
12555 : use pageserver_api::models::TimelineVisibilityState;
12556 :
12557 : use crate::tenant::size::gather_inputs;
12558 :
12559 1 : let tenant_conf = pageserver_api::models::TenantConfig {
12560 1 : // Ensure that we don't compute gc_cutoffs (which needs reading the layer files)
12561 1 : pitr_interval: Some(Duration::ZERO),
12562 1 : ..Default::default()
12563 1 : };
12564 1 : let harness = TenantHarness::create_custom(
12565 1 : "test_synthetic_size_calculation_with_invisible_branches",
12566 1 : tenant_conf,
12567 1 : TenantId::generate(),
12568 1 : ShardIdentity::unsharded(),
12569 1 : Generation::new(0xdeadbeef),
12570 1 : )
12571 1 : .await?;
12572 1 : let (tenant, ctx) = harness.load().await;
12573 1 : let main_tline = tenant
12574 1 : .create_test_timeline_with_layers(
12575 1 : TIMELINE_ID,
12576 1 : Lsn(0x10),
12577 1 : DEFAULT_PG_VERSION,
12578 1 : &ctx,
12579 1 : vec![],
12580 1 : vec![],
12581 1 : vec![],
12582 1 : Lsn(0x100),
12583 1 : )
12584 1 : .await?;
12585 :
12586 1 : let snapshot1 = TimelineId::from_array(hex!("11223344556677881122334455667790"));
12587 1 : tenant
12588 1 : .branch_timeline_test_with_layers(
12589 1 : &main_tline,
12590 1 : snapshot1,
12591 1 : Some(Lsn(0x20)),
12592 1 : &ctx,
12593 1 : vec![],
12594 1 : vec![],
12595 1 : Lsn(0x50),
12596 1 : )
12597 1 : .await?;
12598 1 : let snapshot2 = TimelineId::from_array(hex!("11223344556677881122334455667791"));
12599 1 : tenant
12600 1 : .branch_timeline_test_with_layers(
12601 1 : &main_tline,
12602 1 : snapshot2,
12603 1 : Some(Lsn(0x30)),
12604 1 : &ctx,
12605 1 : vec![],
12606 1 : vec![],
12607 1 : Lsn(0x50),
12608 1 : )
12609 1 : .await?;
12610 1 : let snapshot3 = TimelineId::from_array(hex!("11223344556677881122334455667792"));
12611 1 : tenant
12612 1 : .branch_timeline_test_with_layers(
12613 1 : &main_tline,
12614 1 : snapshot3,
12615 1 : Some(Lsn(0x40)),
12616 1 : &ctx,
12617 1 : vec![],
12618 1 : vec![],
12619 1 : Lsn(0x50),
12620 1 : )
12621 1 : .await?;
12622 1 : let limit = Arc::new(Semaphore::new(1));
12623 1 : let max_retention_period = None;
12624 1 : let mut logical_size_cache = HashMap::new();
12625 1 : let cause = LogicalSizeCalculationCause::EvictionTaskImitation;
12626 1 : let cancel = CancellationToken::new();
12627 :
12628 1 : let inputs = gather_inputs(
12629 1 : &tenant,
12630 1 : &limit,
12631 1 : max_retention_period,
12632 1 : &mut logical_size_cache,
12633 1 : cause,
12634 1 : &cancel,
12635 1 : &ctx,
12636 : )
12637 1 : .instrument(info_span!(
12638 : "gather_inputs",
12639 : tenant_id = "unknown",
12640 : shard_id = "unknown",
12641 : ))
12642 1 : .await?;
12643 : use crate::tenant::size::{LsnKind, ModelInputs, SegmentMeta};
12644 : use LsnKind::*;
12645 : use tenant_size_model::Segment;
12646 1 : let ModelInputs { mut segments, .. } = inputs;
12647 15 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12648 6 : for segment in segments.iter_mut() {
12649 6 : segment.segment.parent = None; // We don't care about the parent for the test
12650 6 : segment.segment.size = None; // We don't care about the size for the test
12651 6 : }
12652 1 : assert_eq!(
12653 : segments,
12654 : [
12655 : SegmentMeta {
12656 : segment: Segment {
12657 : parent: None,
12658 : lsn: 0x10,
12659 : size: None,
12660 : needed: false,
12661 : },
12662 : timeline_id: TIMELINE_ID,
12663 : kind: BranchStart,
12664 : },
12665 : SegmentMeta {
12666 : segment: Segment {
12667 : parent: None,
12668 : lsn: 0x20,
12669 : size: None,
12670 : needed: false,
12671 : },
12672 : timeline_id: TIMELINE_ID,
12673 : kind: BranchPoint,
12674 : },
12675 : SegmentMeta {
12676 : segment: Segment {
12677 : parent: None,
12678 : lsn: 0x30,
12679 : size: None,
12680 : needed: false,
12681 : },
12682 : timeline_id: TIMELINE_ID,
12683 : kind: BranchPoint,
12684 : },
12685 : SegmentMeta {
12686 : segment: Segment {
12687 : parent: None,
12688 : lsn: 0x40,
12689 : size: None,
12690 : needed: false,
12691 : },
12692 : timeline_id: TIMELINE_ID,
12693 : kind: BranchPoint,
12694 : },
12695 : SegmentMeta {
12696 : segment: Segment {
12697 : parent: None,
12698 : lsn: 0x100,
12699 : size: None,
12700 : needed: false,
12701 : },
12702 : timeline_id: TIMELINE_ID,
12703 : kind: GcCutOff,
12704 : }, // we need to retain everything above the last branch point
12705 : SegmentMeta {
12706 : segment: Segment {
12707 : parent: None,
12708 : lsn: 0x100,
12709 : size: None,
12710 : needed: true,
12711 : },
12712 : timeline_id: TIMELINE_ID,
12713 : kind: BranchEnd,
12714 : },
12715 : ]
12716 : );
12717 :
12718 1 : main_tline
12719 1 : .remote_client
12720 1 : .schedule_index_upload_for_timeline_invisible_state(
12721 1 : TimelineVisibilityState::Invisible,
12722 0 : )?;
12723 1 : main_tline.remote_client.wait_completion().await?;
12724 1 : let inputs = gather_inputs(
12725 1 : &tenant,
12726 1 : &limit,
12727 1 : max_retention_period,
12728 1 : &mut logical_size_cache,
12729 1 : cause,
12730 1 : &cancel,
12731 1 : &ctx,
12732 : )
12733 1 : .instrument(info_span!(
12734 : "gather_inputs",
12735 : tenant_id = "unknown",
12736 : shard_id = "unknown",
12737 : ))
12738 1 : .await?;
12739 1 : let ModelInputs { mut segments, .. } = inputs;
12740 14 : segments.retain(|s| s.timeline_id == TIMELINE_ID);
12741 5 : for segment in segments.iter_mut() {
12742 5 : segment.segment.parent = None; // We don't care about the parent for the test
12743 5 : segment.segment.size = None; // We don't care about the size for the test
12744 5 : }
12745 1 : assert_eq!(
12746 : segments,
12747 : [
12748 : SegmentMeta {
12749 : segment: Segment {
12750 : parent: None,
12751 : lsn: 0x10,
12752 : size: None,
12753 : needed: false,
12754 : },
12755 : timeline_id: TIMELINE_ID,
12756 : kind: BranchStart,
12757 : },
12758 : SegmentMeta {
12759 : segment: Segment {
12760 : parent: None,
12761 : lsn: 0x20,
12762 : size: None,
12763 : needed: false,
12764 : },
12765 : timeline_id: TIMELINE_ID,
12766 : kind: BranchPoint,
12767 : },
12768 : SegmentMeta {
12769 : segment: Segment {
12770 : parent: None,
12771 : lsn: 0x30,
12772 : size: None,
12773 : needed: false,
12774 : },
12775 : timeline_id: TIMELINE_ID,
12776 : kind: BranchPoint,
12777 : },
12778 : SegmentMeta {
12779 : segment: Segment {
12780 : parent: None,
12781 : lsn: 0x40,
12782 : size: None,
12783 : needed: false,
12784 : },
12785 : timeline_id: TIMELINE_ID,
12786 : kind: BranchPoint,
12787 : },
12788 : SegmentMeta {
12789 : segment: Segment {
12790 : parent: None,
12791 : lsn: 0x40, // Branch end LSN == last branch point LSN
12792 : size: None,
12793 : needed: true,
12794 : },
12795 : timeline_id: TIMELINE_ID,
12796 : kind: BranchEnd,
12797 : },
12798 : ]
12799 : );
12800 2 : Ok(())
12801 1 : }
12802 : }
|
